sin/samba-mirror
1
0
Fork 0
mirror of https://github.com/samba-team/samba.git synced 2024-12-24 21:34:56 +03:00
Code
68b2e30ae6
samba-mirror/lib/ntdb/hash.c

625 lines
16 KiB
C
Raw Normal View History

TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
/*
Trivial Database 2: hash handling
Copyright (C) Rusty Russell 2010
This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 3 of the License, or (at your option) any later version.
This library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with this library; if not, see <http://www.gnu.org/licenses/>.
*/
#include "private.h"
#include <ccan/hash/hash.h>
/* Default hash function. */
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
uint32_t ntdb_jenkins_hash(const void *key, size_t length, uint32_t seed,
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
void *unused)
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return hash_stable((const unsigned char *)key, length, seed);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
uint32_t ntdb_hash(struct ntdb_context *ntdb, const void *ptr, size_t len)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
return ntdb->hash_fn(ptr, len, ntdb->hash_seed, ntdb->hash_data);
}
static ntdb_bool_err key_matches(struct ntdb_context *ntdb,
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
const struct ntdb_used_record *rec,
ntdb_off_t off,
ntdb: optimize ntdb_fetch. We access the key on lookup, then access the data in the caller. It makes more sense to access both at once. We also put in a likely() for the case where the hash is not chained. Before: Adding 1000 records: 3644-3724(3675) ns (129656 bytes) Finding 1000 records: 1596-1696(1622) ns (129656 bytes) Missing 1000 records: 1409-1525(1452) ns (129656 bytes) Traversing 1000 records: 1636-1747(1668) ns (129656 bytes) Deleting 1000 records: 3138-3223(3175) ns (129656 bytes) Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes) Appending 1000 records: 5396-5529(5426) ns (253312 bytes) Churning 1000 records: 9451-10095(9584) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 183881-191112(188223) ops/sec After: Adding 1000 records: 3590-3701(3640) ns (129656 bytes) Finding 1000 records: 1539-1605(1566) ns (129656 bytes) Missing 1000 records: 1398-1440(1413) ns (129656 bytes) Traversing 1000 records: 1629-2015(1710) ns (129656 bytes) Deleting 1000 records: 3118-3236(3163) ns (129656 bytes) Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes) Appending 1000 records: 5335-5444(5385) ns (253312 bytes) Churning 1000 records: 9350-9955(9494) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 180559-199981(195106) ops/sec
2012-06-19 07:13:09 +04:00
const NTDB_DATA *key,
const char **rptr)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb_bool_err ret = false;
const char *rkey;
if (rec_key_length(rec) != key->dsize) {
ntdb->stats.compare_wrong_keylen++;
return ret;
}
ntdb: optimize ntdb_fetch. We access the key on lookup, then access the data in the caller. It makes more sense to access both at once. We also put in a likely() for the case where the hash is not chained. Before: Adding 1000 records: 3644-3724(3675) ns (129656 bytes) Finding 1000 records: 1596-1696(1622) ns (129656 bytes) Missing 1000 records: 1409-1525(1452) ns (129656 bytes) Traversing 1000 records: 1636-1747(1668) ns (129656 bytes) Deleting 1000 records: 3138-3223(3175) ns (129656 bytes) Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes) Appending 1000 records: 5396-5529(5426) ns (253312 bytes) Churning 1000 records: 9451-10095(9584) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 183881-191112(188223) ops/sec After: Adding 1000 records: 3590-3701(3640) ns (129656 bytes) Finding 1000 records: 1539-1605(1566) ns (129656 bytes) Missing 1000 records: 1398-1440(1413) ns (129656 bytes) Traversing 1000 records: 1629-2015(1710) ns (129656 bytes) Deleting 1000 records: 3118-3236(3163) ns (129656 bytes) Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes) Appending 1000 records: 5335-5444(5385) ns (253312 bytes) Churning 1000 records: 9350-9955(9494) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 180559-199981(195106) ops/sec
2012-06-19 07:13:09 +04:00
rkey = ntdb_access_read(ntdb, off + sizeof(*rec),
key->dsize + rec_data_length(rec), false);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
if (NTDB_PTR_IS_ERR(rkey)) {
return (ntdb_bool_err)NTDB_PTR_ERR(rkey);
}
ntdb: optimize ntdb_fetch. We access the key on lookup, then access the data in the caller. It makes more sense to access both at once. We also put in a likely() for the case where the hash is not chained. Before: Adding 1000 records: 3644-3724(3675) ns (129656 bytes) Finding 1000 records: 1596-1696(1622) ns (129656 bytes) Missing 1000 records: 1409-1525(1452) ns (129656 bytes) Traversing 1000 records: 1636-1747(1668) ns (129656 bytes) Deleting 1000 records: 3138-3223(3175) ns (129656 bytes) Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes) Appending 1000 records: 5396-5529(5426) ns (253312 bytes) Churning 1000 records: 9451-10095(9584) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 183881-191112(188223) ops/sec After: Adding 1000 records: 3590-3701(3640) ns (129656 bytes) Finding 1000 records: 1539-1605(1566) ns (129656 bytes) Missing 1000 records: 1398-1440(1413) ns (129656 bytes) Traversing 1000 records: 1629-2015(1710) ns (129656 bytes) Deleting 1000 records: 3118-3236(3163) ns (129656 bytes) Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes) Appending 1000 records: 5335-5444(5385) ns (253312 bytes) Churning 1000 records: 9350-9955(9494) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 180559-199981(195106) ops/sec
2012-06-19 07:13:09 +04:00
if (memcmp(rkey, key->dptr, key->dsize) == 0) {
if (rptr) {
*rptr = rkey;
} else {
ntdb_access_release(ntdb, rkey);
}
return true;
}
ntdb->stats.compare_wrong_keycmp++;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb_access_release(ntdb, rkey);
return ret;
}
/* Does entry match? */
static ntdb_bool_err match(struct ntdb_context *ntdb,
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
uint32_t hash,
const NTDB_DATA *key,
ntdb_off_t val,
struct ntdb_used_record *rec,
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
const char **rptr)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb_off_t off;
enum NTDB_ERROR ecode;
ntdb->stats.compares++;
/* Top bits of offset == next bits of hash. */
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (bits_from(hash, ntdb->hash_bits, NTDB_OFF_UPPER_STEAL)
!= bits_from(val, 64-NTDB_OFF_UPPER_STEAL, NTDB_OFF_UPPER_STEAL)) {
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb->stats.compare_wrong_offsetbits++;
return false;
}
off = val & NTDB_OFF_MASK;
ecode = ntdb_read_convert(ntdb, off, rec, sizeof(*rec));
if (ecode != NTDB_SUCCESS) {
return (ntdb_bool_err)ecode;
}
ntdb: optimize ntdb_fetch. We access the key on lookup, then access the data in the caller. It makes more sense to access both at once. We also put in a likely() for the case where the hash is not chained. Before: Adding 1000 records: 3644-3724(3675) ns (129656 bytes) Finding 1000 records: 1596-1696(1622) ns (129656 bytes) Missing 1000 records: 1409-1525(1452) ns (129656 bytes) Traversing 1000 records: 1636-1747(1668) ns (129656 bytes) Deleting 1000 records: 3138-3223(3175) ns (129656 bytes) Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes) Appending 1000 records: 5396-5529(5426) ns (253312 bytes) Churning 1000 records: 9451-10095(9584) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 183881-191112(188223) ops/sec After: Adding 1000 records: 3590-3701(3640) ns (129656 bytes) Finding 1000 records: 1539-1605(1566) ns (129656 bytes) Missing 1000 records: 1398-1440(1413) ns (129656 bytes) Traversing 1000 records: 1629-2015(1710) ns (129656 bytes) Deleting 1000 records: 3118-3236(3163) ns (129656 bytes) Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes) Appending 1000 records: 5335-5444(5385) ns (253312 bytes) Churning 1000 records: 9350-9955(9494) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 180559-199981(195106) ops/sec
2012-06-19 07:13:09 +04:00
return key_matches(ntdb, rec, off, key, rptr);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
static bool is_chain(ntdb_off_t val)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return val & (1ULL << NTDB_OFF_CHAIN_BIT);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
static ntdb_off_t hbucket_off(ntdb_off_t base, ntdb_len_t idx)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return base + sizeof(struct ntdb_used_record)
+ idx * sizeof(ntdb_off_t);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
/* This is the core routine which searches the hashtable for an entry.
* On error, no locks are held and -ve is returned.
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
* Otherwise, hinfo is filled in.
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
* If not found, the return value is 0.
* If found, the return value is the offset, and *rec is the record. */
ntdb_off_t find_and_lock(struct ntdb_context *ntdb,
ntdb: optimize ntdb_fetch. We access the key on lookup, then access the data in the caller. It makes more sense to access both at once. We also put in a likely() for the case where the hash is not chained. Before: Adding 1000 records: 3644-3724(3675) ns (129656 bytes) Finding 1000 records: 1596-1696(1622) ns (129656 bytes) Missing 1000 records: 1409-1525(1452) ns (129656 bytes) Traversing 1000 records: 1636-1747(1668) ns (129656 bytes) Deleting 1000 records: 3138-3223(3175) ns (129656 bytes) Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes) Appending 1000 records: 5396-5529(5426) ns (253312 bytes) Churning 1000 records: 9451-10095(9584) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 183881-191112(188223) ops/sec After: Adding 1000 records: 3590-3701(3640) ns (129656 bytes) Finding 1000 records: 1539-1605(1566) ns (129656 bytes) Missing 1000 records: 1398-1440(1413) ns (129656 bytes) Traversing 1000 records: 1629-2015(1710) ns (129656 bytes) Deleting 1000 records: 3118-3236(3163) ns (129656 bytes) Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes) Appending 1000 records: 5335-5444(5385) ns (253312 bytes) Churning 1000 records: 9350-9955(9494) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 180559-199981(195106) ops/sec
2012-06-19 07:13:09 +04:00
NTDB_DATA key,
int ltype,
struct hash_info *h,
struct ntdb_used_record *rec,
const char **rptr)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_off_t off, val;
const ntdb_off_t *arr = NULL;
ntdb_len_t i;
bool found_empty;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
enum NTDB_ERROR ecode;
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
struct ntdb_used_record chdr;
ntdb_bool_err berr;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
h->h = ntdb_hash(ntdb, key.dptr, key.dsize);
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
h->table = NTDB_HASH_OFFSET;
h->table_size = 1 << ntdb->hash_bits;
h->bucket = bits_from(h->h, 0, ntdb->hash_bits);
h->old_val = 0;
ecode = ntdb_lock_hash(ntdb, h->bucket, ltype);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
if (ecode != NTDB_SUCCESS) {
return NTDB_ERR_TO_OFF(ecode);
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
off = hbucket_off(h->table, h->bucket);
val = ntdb_read_off(ntdb, off);
if (NTDB_OFF_IS_ERR(val)) {
ecode = NTDB_OFF_TO_ERR(val);
goto fail;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Directly in hash table? */
ntdb: optimize ntdb_fetch. We access the key on lookup, then access the data in the caller. It makes more sense to access both at once. We also put in a likely() for the case where the hash is not chained. Before: Adding 1000 records: 3644-3724(3675) ns (129656 bytes) Finding 1000 records: 1596-1696(1622) ns (129656 bytes) Missing 1000 records: 1409-1525(1452) ns (129656 bytes) Traversing 1000 records: 1636-1747(1668) ns (129656 bytes) Deleting 1000 records: 3138-3223(3175) ns (129656 bytes) Re-adding 1000 records: 3278-3414(3329) ns (129656 bytes) Appending 1000 records: 5396-5529(5426) ns (253312 bytes) Churning 1000 records: 9451-10095(9584) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 183881-191112(188223) ops/sec After: Adding 1000 records: 3590-3701(3640) ns (129656 bytes) Finding 1000 records: 1539-1605(1566) ns (129656 bytes) Missing 1000 records: 1398-1440(1413) ns (129656 bytes) Traversing 1000 records: 1629-2015(1710) ns (129656 bytes) Deleting 1000 records: 3118-3236(3163) ns (129656 bytes) Re-adding 1000 records: 3235-3355(3275) ns (129656 bytes) Appending 1000 records: 5335-5444(5385) ns (253312 bytes) Churning 1000 records: 9350-9955(9494) ns (253312 bytes) smbtorture results (--entries=1000) ntdb speed 180559-199981(195106) ops/sec
2012-06-19 07:13:09 +04:00
if (!likely(is_chain(val))) {
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (val) {
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
berr = match(ntdb, h->h, &key, val, rec, rptr);
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (berr < 0) {
ecode = NTDB_OFF_TO_ERR(berr);
goto fail;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (berr) {
return val & NTDB_OFF_MASK;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* If you want to insert here, make a chain. */
h->old_val = val;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return 0;
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Nope? Iterate through chain. */
h->table = val & NTDB_OFF_MASK;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ecode = ntdb_read_convert(ntdb, h->table, &chdr, sizeof(chdr));
if (ecode != NTDB_SUCCESS) {
goto fail;
}
if (rec_magic(&chdr) != NTDB_CHAIN_MAGIC) {
ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
NTDB_LOG_ERROR,
"find_and_lock:"
" corrupt record %#x at %llu",
rec_magic(&chdr), (long long)off);
goto fail;
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
h->table_size = rec_data_length(&chdr) / sizeof(ntdb_off_t);
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
arr = ntdb_access_read(ntdb, hbucket_off(h->table, 0),
rec_data_length(&chdr), true);
if (NTDB_PTR_IS_ERR(arr)) {
ecode = NTDB_PTR_ERR(arr);
goto fail;
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
found_empty = false;
for (i = 0; i < h->table_size; i++) {
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
if (arr[i] == 0) {
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (!found_empty) {
h->bucket = i;
found_empty = true;
}
} else {
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
berr = match(ntdb, h->h, &key, arr[i], rec, rptr);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
if (berr < 0) {
ecode = NTDB_OFF_TO_ERR(berr);
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
ntdb_access_release(ntdb, arr);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
goto fail;
}
if (berr) {
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* We found it! */
h->bucket = i;
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
off = arr[i] & NTDB_OFF_MASK;
ntdb_access_release(ntdb, arr);
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return off;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
}
if (!found_empty) {
/* Set to any non-zero value */
h->old_val = 1;
h->bucket = i;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: take advantage of direct access across expand. This means we no longer have to unmap if we want to compare a record. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-22 04:14:41 +04:00
ntdb_access_release(ntdb, arr);
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return 0;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
fail:
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_unlock_hash(ntdb, h->bucket, ltype);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
return NTDB_ERR_TO_OFF(ecode);
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
static ntdb_off_t encode_offset(const struct ntdb_context *ntdb,
ntdb_off_t new_off, uint32_t hash)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_off_t extra;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
assert((new_off & (1ULL << NTDB_OFF_CHAIN_BIT)) == 0);
assert((new_off >> (64 - NTDB_OFF_UPPER_STEAL)) == 0);
/* We pack extra hash bits into the upper bits of the offset. */
extra = bits_from(hash, ntdb->hash_bits, NTDB_OFF_UPPER_STEAL);
extra <<= (64 - NTDB_OFF_UPPER_STEAL);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return new_off | extra;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Simply overwrite the hash entry we found before. */
enum NTDB_ERROR replace_in_hash(struct ntdb_context *ntdb,
const struct hash_info *h,
ntdb_off_t new_off)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return ntdb_write_off(ntdb, hbucket_off(h->table, h->bucket),
encode_offset(ntdb, new_off, h->h));
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
enum NTDB_ERROR delete_from_hash(struct ntdb_context *ntdb,
const struct hash_info *h)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return ntdb_write_off(ntdb, hbucket_off(h->table, h->bucket), 0);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
enum NTDB_ERROR add_to_hash(struct ntdb_context *ntdb,
const struct hash_info *h,
ntdb_off_t new_off)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
enum NTDB_ERROR ecode;
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_off_t chain;
struct ntdb_used_record chdr;
const ntdb_off_t *old;
ntdb_off_t *new;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* We hit an empty bucket during search? That's where it goes. */
if (!h->old_val) {
return replace_in_hash(ntdb, h, new_off);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Full at top-level? Create a 2-element chain. */
if (h->table == NTDB_HASH_OFFSET) {
ntdb_off_t pair[2];
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* One element is old value, the other is the new value. */
pair[0] = h->old_val;
pair[1] = encode_offset(ntdb, new_off, h->h);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
chain = alloc(ntdb, 0, sizeof(pair), NTDB_CHAIN_MAGIC, true);
if (NTDB_OFF_IS_ERR(chain)) {
return NTDB_OFF_TO_ERR(chain);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ecode = ntdb_write_convert(ntdb,
chain
+ sizeof(struct ntdb_used_record),
pair, sizeof(pair));
if (ecode == NTDB_SUCCESS) {
ecode = ntdb_write_off(ntdb,
hbucket_off(h->table, h->bucket),
chain
| (1ULL << NTDB_OFF_CHAIN_BIT));
}
return ecode;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Full bucket. Expand. */
ecode = ntdb_read_convert(ntdb, h->table, &chdr, sizeof(chdr));
if (ecode != NTDB_SUCCESS) {
return ecode;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (rec_extra_padding(&chdr) >= sizeof(new_off)) {
/* Expand in place. */
uint64_t dlen = rec_data_length(&chdr);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ecode = set_header(ntdb, &chdr, NTDB_CHAIN_MAGIC, 0,
dlen + sizeof(new_off),
dlen + rec_extra_padding(&chdr));
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (ecode != NTDB_SUCCESS) {
return ecode;
}
/* find_and_lock set up h to point to last bucket. */
ecode = replace_in_hash(ntdb, h, new_off);
if (ecode != NTDB_SUCCESS) {
return ecode;
}
ecode = ntdb_write_convert(ntdb, h->table, &chdr, sizeof(chdr));
if (ecode != NTDB_SUCCESS) {
return ecode;
}
/* For futureproofing, we always make the first byte of padding
* a zero. */
if (rec_extra_padding(&chdr)) {
ecode = ntdb->io->twrite(ntdb, h->table + sizeof(chdr)
+ dlen + sizeof(new_off),
"", 1);
}
return ecode;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* We need to reallocate the chain. */
chain = alloc(ntdb, 0, (h->table_size + 1) * sizeof(ntdb_off_t),
NTDB_CHAIN_MAGIC, true);
if (NTDB_OFF_IS_ERR(chain)) {
return NTDB_OFF_TO_ERR(chain);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Map both and copy across old buckets. */
old = ntdb_access_read(ntdb, hbucket_off(h->table, 0),
h->table_size*sizeof(ntdb_off_t), true);
if (NTDB_PTR_IS_ERR(old)) {
return NTDB_PTR_ERR(old);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
new = ntdb_access_write(ntdb, hbucket_off(chain, 0),
(h->table_size + 1)*sizeof(ntdb_off_t), true);
if (NTDB_PTR_IS_ERR(new)) {
ntdb_access_release(ntdb, old);
return NTDB_PTR_ERR(new);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
memcpy(new, old, h->bucket * sizeof(ntdb_off_t));
new[h->bucket] = encode_offset(ntdb, new_off, h->h);
ntdb_access_release(ntdb, old);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ecode = ntdb_access_commit(ntdb, new);
if (ecode != NTDB_SUCCESS) {
return ecode;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Free the old chain. */
ecode = add_free_record(ntdb, h->table,
sizeof(struct ntdb_used_record)
+ rec_data_length(&chdr)
+ rec_extra_padding(&chdr),
NTDB_LOCK_WAIT, true);
/* Replace top-level to point to new chain */
return ntdb_write_off(ntdb,
hbucket_off(NTDB_HASH_OFFSET,
bits_from(h->h, 0, ntdb->hash_bits)),
chain | (1ULL << NTDB_OFF_CHAIN_BIT));
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Traverse support: returns offset of record, or 0 or -ve error. */
static ntdb_off_t iterate_chain(struct ntdb_context *ntdb,
ntdb_off_t val,
struct hash_info *h)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_off_t i;
enum NTDB_ERROR ecode;
struct ntdb_used_record chdr;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* First load up chain header. */
h->table = val & NTDB_OFF_MASK;
ecode = ntdb_read_convert(ntdb, h->table, &chdr, sizeof(chdr));
if (ecode != NTDB_SUCCESS) {
return ecode;
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (rec_magic(&chdr) != NTDB_CHAIN_MAGIC) {
return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
NTDB_LOG_ERROR,
"get_table:"
" corrupt record %#x at %llu",
rec_magic(&chdr),
(long long)h->table);
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Chain length is implied by data length. */
h->table_size = rec_data_length(&chdr) / sizeof(ntdb_off_t);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
i = ntdb_find_nonzero_off(ntdb, hbucket_off(h->table, 0), h->bucket,
h->table_size);
if (NTDB_OFF_IS_ERR(i)) {
return i;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (i != h->table_size) {
/* Return to next bucket. */
h->bucket = i + 1;
val = ntdb_read_off(ntdb, hbucket_off(h->table, i));
if (NTDB_OFF_IS_ERR(val)) {
return val;
}
return val & NTDB_OFF_MASK;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Go back up to hash table. */
h->table = NTDB_HASH_OFFSET;
h->table_size = 1 << ntdb->hash_bits;
h->bucket = bits_from(h->h, 0, ntdb->hash_bits) + 1;
return 0;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Keeps hash locked unless returns 0 or error. */
static ntdb_off_t lock_and_iterate_hash(struct ntdb_context *ntdb,
struct hash_info *h)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_off_t val, i;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
enum NTDB_ERROR ecode;
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (h->table != NTDB_HASH_OFFSET) {
/* We're in a chain. */
i = bits_from(h->h, 0, ntdb->hash_bits);
ecode = ntdb_lock_hash(ntdb, i, F_RDLCK);
if (ecode != NTDB_SUCCESS) {
return NTDB_ERR_TO_OFF(ecode);
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* We dropped lock, bucket might have moved! */
val = ntdb_read_off(ntdb, hbucket_off(NTDB_HASH_OFFSET, i));
if (NTDB_OFF_IS_ERR(val)) {
goto unlock;
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* We don't remove chains: there should still be one there! */
if (!val || !is_chain(val)) {
ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
NTDB_LOG_ERROR,
"iterate_hash:"
" vanished hchain %llu at %llu",
(long long)val,
(long long)i);
val = NTDB_ERR_TO_OFF(ecode);
goto unlock;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Find next bucket in the chain. */
val = iterate_chain(ntdb, val, h);
if (NTDB_OFF_IS_ERR(val)) {
goto unlock;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (val != 0) {
return val;
}
ntdb_unlock_hash(ntdb, i, F_RDLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* OK, we've reset h back to top level. */
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* We do this unlocked, then re-check. */
for (i = ntdb_find_nonzero_off(ntdb, hbucket_off(h->table, 0),
h->bucket, h->table_size);
i != h->table_size;
i = ntdb_find_nonzero_off(ntdb, hbucket_off(h->table, 0),
i+1, h->table_size)) {
ecode = ntdb_lock_hash(ntdb, i, F_RDLCK);
if (ecode != NTDB_SUCCESS) {
return NTDB_ERR_TO_OFF(ecode);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
val = ntdb_read_off(ntdb, hbucket_off(h->table, i));
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
if (NTDB_OFF_IS_ERR(val)) {
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
goto unlock;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Lost race, and it's empty? */
if (!val) {
ntdb->stats.traverse_val_vanished++;
ntdb_unlock_hash(ntdb, i, F_RDLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
continue;
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (!is_chain(val)) {
/* So caller knows what lock to free. */
h->h = i;
/* Return to next bucket. */
h->bucket = i + 1;
val &= NTDB_OFF_MASK;
return val;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Start at beginning of chain */
h->bucket = 0;
h->h = i;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
val = iterate_chain(ntdb, val, h);
if (NTDB_OFF_IS_ERR(val)) {
goto unlock;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (val != 0) {
return val;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* Otherwise, bucket has been set to i+1 */
ntdb_unlock_hash(ntdb, i, F_RDLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return 0;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
unlock:
ntdb_unlock_hash(ntdb, i, F_RDLCK);
return val;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
/* Return success if we find something, NTDB_ERR_NOEXIST if none. */
enum NTDB_ERROR next_in_hash(struct ntdb_context *ntdb,
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
struct hash_info *h,
NTDB_DATA *kbuf, size_t *dlen)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_off_t off;
struct ntdb_used_record rec;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
enum NTDB_ERROR ecode;
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
off = lock_and_iterate_hash(ntdb, h);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (NTDB_OFF_IS_ERR(off)) {
return NTDB_OFF_TO_ERR(off);
} else if (off == 0) {
return NTDB_ERR_NOEXIST;
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* The hash for this key is still locked. */
ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
if (ecode != NTDB_SUCCESS) {
goto unlock;
}
if (rec_magic(&rec) != NTDB_USED_MAGIC) {
ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT,
NTDB_LOG_ERROR,
"next_in_hash:"
" corrupt record at %llu",
(long long)off);
goto unlock;
}
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
kbuf->dsize = rec_key_length(&rec);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
/* They want data as well? */
if (dlen) {
*dlen = rec_data_length(&rec);
kbuf->dptr = ntdb_alloc_read(ntdb, off + sizeof(rec),
kbuf->dsize + *dlen);
} else {
kbuf->dptr = ntdb_alloc_read(ntdb, off + sizeof(rec),
kbuf->dsize);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
if (NTDB_PTR_IS_ERR(kbuf->dptr)) {
ecode = NTDB_PTR_ERR(kbuf->dptr);
goto unlock;
}
ecode = NTDB_SUCCESS;
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
unlock:
ntdb_unlock_hash(ntdb, bits_from(h->h, 0, ntdb->hash_bits), F_RDLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
return ecode;
}
enum NTDB_ERROR first_in_hash(struct ntdb_context *ntdb,
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
struct hash_info *h,
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
NTDB_DATA *kbuf, size_t *dlen)
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
h->table = NTDB_HASH_OFFSET;
h->table_size = 1 << ntdb->hash_bits;
h->bucket = 0;
return next_in_hash(ntdb, h, kbuf, dlen);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
/* Even if the entry isn't in this hash bucket, you'd have to lock this
* bucket to find it. */
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
static enum NTDB_ERROR chainlock(struct ntdb_context *ntdb,
const NTDB_DATA *key, int ltype)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
uint32_t h = ntdb_hash(ntdb, key->dptr, key->dsize);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return ntdb_lock_hash(ntdb, bits_from(h, 0, ntdb->hash_bits), ltype);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
/* lock/unlock one hash chain. This is meant to be used to reduce
contention - it cannot guarantee how many records will be locked */
_PUBLIC_ enum NTDB_ERROR ntdb_chainlock(struct ntdb_context *ntdb, NTDB_DATA key)
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return chainlock(ntdb, &key, F_WRLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
_PUBLIC_ void ntdb_chainunlock(struct ntdb_context *ntdb, NTDB_DATA key)
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
uint32_t h = ntdb_hash(ntdb, key.dptr, key.dsize);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_unlock_hash(ntdb, bits_from(h, 0, ntdb->hash_bits), F_WRLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
_PUBLIC_ enum NTDB_ERROR ntdb_chainlock_read(struct ntdb_context *ntdb,
NTDB_DATA key)
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
return chainlock(ntdb, &key, F_RDLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
_PUBLIC_ void ntdb_chainunlock_read(struct ntdb_context *ntdb, NTDB_DATA key)
{
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
uint32_t h = ntdb_hash(ntdb, key.dptr, key.dsize);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
ntdb: remove hash table trees. TDB2 started with a top-level hash of 1024 entries, divided into 128 groups of 8 buckets. When a bucket filled, the 8 bucket group expanded into pointers into 8 new 64-entry hash tables. When these filled, they expanded in turn, etc. It's a nice idea to automatically expand the hash tables, but it doesn't pay off. Remove it for NTDB. 1) It only beats TDB performance when the database is huge and the TDB hashsize is small. We are about 20% slower on medium-size databases (1000 to 10000 records), worse on really small ones. 2) Since we're 64 bits, our hash tables are already twice as expensive as TDB. 3) Since our hash function is good, it means that all groups tend to fill at the same time, meaning the hash enlarges by a factor of 128 all at once, leading to a very large database at that point. 4) Our efficiency would improve if we enlarged the top level, but that makes our minimum db size even worse: it's already over 8k, and jumps to 1M after about 1000 entries! 5) Making the sub group size larger gives a shallower tree, which performs better, but makes the "hash explosion" problem worse. 6) The code is complicated, having to handle delete and reshuffling groups of hash buckets, and expansion of buckets. 7) We have to handle the case where all the records somehow end up with the same hash value, which requires special code to chain records for that case. On the other hand, it would be nice if we didn't degrade as badly as TDB does when the hash chains get long. This patch removes the hash-growing code, but instead of chaining like TDB does when a bucket fills, we point the bucket to an array of record pointers. Since each on-disk NTDB pointer contains some hash bits from the record (we steal the upper 8 bits of the offset), 99.5% of the time we don't need to load the record to determine if it matches. This makes an array of offsets much more cache-friendly than a linked list. Here are the times (in ns) for tdb_store of N records, tdb_store of N records the second time, and a fetch of all N records. I've also included the final database size and the smbtorture local.[n]tdb_speed results. Benchmark details: 1) Compiled with -O2. 2) assert() was disabled in TDB2 and NTDB. 3) The "optimize fetch" patch was applied to NTDB. 10 runs, using tmpfs (otherwise massive swapping as db hits ~30M, despite plenty of RAM). Insert Re-ins Fetch Size dbspeed (nsec) (nsec) (nsec) (Kb) (ops/sec) TDB (10000 hashsize): 100 records: 3882 3320 1609 53 203204 1000 records: 3651 3281 1571 115 218021 10000 records: 3404 3326 1595 880 202874 100000 records: 4317 3825 2097 8262 126811 1000000 records: 11568 11578 9320 77005 25046 TDB2 (1024 hashsize, expandable): 100 records: 3867 3329 1699 17 187100 1000 records: 4040 3249 1639 154 186255 10000 records: 4143 3300 1695 1226 185110 100000 records: 4481 3425 1800 17848 163483 1000000 records: 4055 3534 1878 106386 160774 NTDB (8192 hashsize) 100 records: 4259 3376 1692 82 190852 1000 records: 3640 3275 1566 130 195106 10000 records: 4337 3438 1614 773 188362 100000 records: 4750 5165 1746 9001 169197 1000000 records: 4897 5180 2341 83838 121901 Analysis: 1) TDB wins on small databases, beating TDB2 by ~15%, NTDB by ~10%. 2) TDB starts to lose when hash chains get 10 long (fetch 10% slower than TDB2/NTDB). 3) TDB does horribly when hash chains get 100 long (fetch 4x slower than NTDB, 5x slower than TDB2, insert about 2-3x slower). 4) TDB2 databases are 40% larger than TDB1. NTDB is about 15% larger than TDB1
2012-06-19 07:13:04 +04:00
ntdb_unlock_hash(ntdb, bits_from(h, 0, ntdb->hash_bits), F_RDLCK);
TDB2: Goodbye TDB2, Hello NTDB. This renames everything from tdb2 to ntdb: importantly, we no longer use the tdb_ namespace, so you can link against both ntdb and tdb if you want to. This also enables building of standalone ntdb by the autobuild script. Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
2012-06-18 17:00:26 +04:00
}
Copy Permalink