1
0
mirror of https://github.com/samba-team/samba.git synced 2024-12-24 21:34:56 +03:00
Commit Graph

87 Commits

Author SHA1 Message Date
Tim Beale
9f504fd5a3 traffic_replay: Store total conversations on the replay context
This is useful info to know, and will be used in subsequent commits.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-07-24 02:24:27 +00:00
Joe Guo
2ee72cc615 traffic: load dns query from file and write stats to file
Signed-off-by: Joe Guo <joeg@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
Reviewed-by: Andreas Schneider <asn@samba.org>

Autobuild-User(master): Andrew Bartlett <abartlet@samba.org>
Autobuild-Date(master): Wed May  1 01:10:42 UTC 2019 on sn-devel-184
2019-05-01 01:10:42 +00:00
Joe Guo
35e52ebdd6 traffic: define kerberos_state to simplify code
Signed-off-by: Joe Guo <joeg@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
Reviewed-by: Andreas Schneider <asn@samba.org>
2019-04-30 23:18:29 +00:00
Joe Guo
e98340e3e2 traffic: make code more pythonic
Signed-off-by: Joe Guo <joeg@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
Reviewed-by: Andreas Schneider <asn@samba.org>
2019-04-30 23:18:28 +00:00
Douglas Bagnall
3c10cecac1 traffic_replay: use packets per second as primary scale
The old -S/--scale-traffic is relative to the original model, which made
its relationship to true traffic volumes quite opaque

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:35 +01:00
Douglas Bagnall
524777e681 traffic: add option to reanimate dying conversations
The traffic model is generated from a window in time, which makes
conversations appear to start and stop unnaturally at the window
boundaries. When the window is short compared to the traffic replay
time and the true expected conversation length, this has a significant
distorting effect, leading to more conversations than would be
expected to generate a given number of packets.

To offset this slightly we add the --conversation-persistence option
which tries to convert apparent death into a longish wait.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:34 +01:00
Douglas Bagnall
b5d493f927 traffic: Conversation.add_short_packet is discerning about packets
If the packets really wouldn't do anything, we might as well not add them.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:34 +01:00
Douglas Bagnall
db51004307 traffic: remove useless branch in stats report
This completes the work of 68c64c634a,
but differs from that in that it makes no actual change because isatty
was not being called so was always evaluated as true.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:34 +01:00
Douglas Bagnall
d727dcc82a traffic: generate more statistics in replay
Add more "header" values indicating the progress of the run as a
whole.

The new fields are:

Max sleep miss            - the longest sleep() oversleep. Indicates client load.
Maximum lag               - the longest gap between a planned packet
                            time and its actual time.
Start lag                 - the longest gap between intended and actual
                            conversation start.
Planned conversations     - how many conversations we meant to have.
Planned packets           - how many "packets" we thought we were making. Not
                            all "packets" result in actual operations or packets.
Unfinished conversations  - how many conversations had not finished
                            when they were killed.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:34 +01:00
Douglas Bagnall
31c0809e4f traffic: avoid bare except: clauses
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:34 +01:00
Douglas Bagnall
41b0cfce2c traffic: simplify tests for badpassword_frequency
x <= 0 will fail one or both of the other test clauses.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:34 +01:00
Douglas Bagnall
65c02fd68c traffic: assign context domain at start
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
17579dd810 traffic: traffic_replay --latency-timeout to control final wait
Conversations that haven't finished within some acceptable margin of
on-time can be said to have failed. This is where you specify that
margin.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
b737552ed3 traffic_replay: --stop-on-any-error option to not ignore client trouble
Sometimes you want to know if any client is crashing for any reason.
In those times use --stop-on-any-error for an early exit.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
22cba7011c traffic: evoke OS error codes by name, not number
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
35bc6ee0a1 traffic: reinitialise RNG in each child
Otherwise they all replay using the same random sequence.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
6268effb11 traffic: avoid generating conversations without packets
Some "packets" don't generate any actual traffic. If we have a
conversation consisting only of those, we can avoid forking a client
for it.

This *slightly* increases the load over that which would be generated
otherwise for a given traffic rate, but that's OK.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
c4d5bb5952 traffic: rework conversation generation to better use memory
Use less memory altogether and don't allocated shared mutable before
the fork.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
7b03e81c61 traffic: generate sparser descriptions of conversations
Rather than building all the packets at this point, we stick to the
barest details of the packets (which is all the model gives us
anyway).

The advantage is that will take a lot less memory, which matters
because this process forks into many clients that were sharing and
mutate the conversation list.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
7edf58dc58 traffic: new version of model with packet_rate, version number
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
273eb3dffb traffic: fix hash non-determinism when loading JSON
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Tim Beale <timbeale@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
ef747953d4 traffic: Conversation gets given id, list of short packets
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:33 +01:00
Douglas Bagnall
a430b11ca2 traffic: rename packet_rate -> replay_speed for accuracy and room
We are soon going to have a self.packet_rate, and replay_speed is more
accurate in this case.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
a352060f97 traffic: initialise conversation specific lists after the fork
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
e866782a15 traffic: split is_a_real_packet() function out of class
So we can use it to determine whether a packet should be a Packet before
making the leap.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
24d09d1df8 traffic: give Packet __slots__ for reduced memory
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
4f7ae5d925 traffic: make random_colour_print more deterministic, silent on demand
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
158f172845 traffic: use namedtuple for ConversationAccounts namespace class
Saving memory, which reduces fork overhead.

Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
9da79b54fa traffic: fix mk_masked_dir doc and comments
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Douglas Bagnall
709d2c84cb traffic: Packet.from_line classmethod uses cls var
Signed-off-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2019-01-08 23:55:32 +01:00
Noel Power
923010d99b various: Remove references to about to be deleted thirdparty/dnspython
Signed-off-by: Noel Power <noel.power@suse.com>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-12-11 20:07:18 +01:00
Tim Beale
5517d60653 traffic_replay: Add a max-members option to cap group size
traffic_replay tries to distribute the users among the groups in a
realistic manner - some groups will have almost all users in them.
However, this becomes a problem when testing a really large database,
e.g. we may want 100K users, but no more than 5K users in each group.

This patch adds a max-member option so we can limit how big the groups
actually get.

If we detect that a group exceeds the max-members, we reset the group's
probability (of getting selected) to zero, and then recalculate the
cumulative distribution. The means that the group should no longer get
selected by generate_random_membership(). (Note we can't completely
remove the group from the list because that changes the
list-index-to-group-ID mapping).

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>

Autobuild-User(master): Andrew Bartlett <abartlet@samba.org>
Autobuild-Date(master): Tue Dec  4 12:22:50 CET 2018 on sn-devel-144
2018-12-04 12:22:50 +01:00
Tim Beale
13f57a7f80 traffic: Rework how assignments are generated slightly
We want to cap the number of members that can be in a group. But first,
we need to tweak how the assignment dict gets generated, so that we get
rid of the intermediary set.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-12-04 09:17:17 +01:00
Tim Beale
a0b5f4b7b0 traffic_replay: Rework machine accounts to remove redundant code
generate_users_and_groups() now generates the machine acounts as well as
the user accounts, so it seems there's no need to also have
generate_traffic_accounts(), which does the same job.

Instead, we can just pass through the number of machine acounts to
generate_users_and_groups() and delete the other function.

Also updated generate_users_and_groups() so that machine_accounts is
no longer optional (we want to create machine accounts in all cases).

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-11-07 17:55:09 +01:00
Tim Beale
1d7fb66f76 traffic_replay: Make packet generation work on a pre-populated DB again
Generate separate machine accounts for populating a large DB vs
replaying network traffic.

We want to use different userAccountControl flags in each of the above
cases (i.e. commit 3338a3e257). However, this means that once you
use the --generate-users-only option, you can't replay network packets
against the machine accounts.

We can avoid this problem by creating separate machine accounts for each
of 2 different cases, e.g. STGM-0-x machines for traffic-replay, and
PC-0-x machines for padding out the database.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-11-07 17:55:09 +01:00
Tim Beale
c7fe481477 traffic_replay: Make sure naming assumptions are in a single place
The traffic_replay group/user/machine account names follow a standard
format. This adds a function to generate the machine-name. It also makes
sure the existing user_name() function gets called in all applicable
places.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-11-07 17:55:09 +01:00
Tim Beale
51917fc07f traffic_replay: Move 'traffic account' flag up a level
We create machine accounts for 2 different purposes:
1). For traffic generation, i.e. testing realistic network packets.
2). For generating a realistic large DB.

Unfortunately, we want to use different userAccountControl flags for
the 2 different cases. Commit 3338a3e257 changed the flags used
for case #2, but this breaks case #1.

The problem is generate_users_and_groups() is called in both cases,
so we want the 'traffic account' flag passed into that function.
This ensures that the machine accounts get created with the appropriate
userAccountControl flags for the particular case you want to test.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-11-07 17:55:08 +01:00
Tim Beale
85b6d88989 traffic_replay: Move machine account creation
I was assuming that generate_users_and_groups() only gets called in the
--generate-users-only case. However, it also gets called in the default
traffic replay case.

This patch reworks the code so that the number of machine accounts to
create gets passed in, and the 'create 25% more computers than users'
assumption only applies to the --generate-users-only case.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Andrew Bartlett <abartlet@samba.org>
2018-11-07 17:55:08 +01:00
Tim Beale
3338a3e257 traffic: Machine accounts were generated as critical objects
Due to the userAccountControl flags we were specifying, the machine
accounts were all created as critical objects. When trying to populate
1000s of machine accounts in a DB, this makes replication unnecessarily
slow (because it has to replicate them all twice).

This patch changes it so when we're just creating machine accounts for
the purpose of populating a semi-realistic DB, we jsut use the default
WORKSTATION_TRUST_ACCOUNT flag.

Note that for the accounts used for traffic-replay, we apparently need
the existing flags in order for the DC to accept certain requests.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>

Autobuild-User(master): Tim Beale <timbeale@samba.org>
Autobuild-Date(master): Mon Nov  5 03:43:24 CET 2018 on sn-devel-144
2018-11-05 03:43:24 +01:00
Tim Beale
be51b51263 traffic_replay: Generate machine accounts as well as users
Currently the tool only generates the machine accounts needed for
traffic generation. However, this isn't realistic if we're trying to use
the tool to generate users to simulate a large network.

This patch generates machine accoutns along with the user accounts.
Note we assume there will be more computer accounts than users in a real
network (e.g. work laptops, servers, etc), so generate slightly more
computer accounts.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:17 +01:00
Tim Beale
1906312c09 traffic_replay: Improve user generation debug
When creating 1000s of users you currently get a lot of debug, but at
the same time you have no idea how far through creating the users you
actually are.

Instead of logging every single user account that's created, log every
50th (as well as how far through the overall generation we are).

Logger already includes timestamps, so we can remove generating the
timestamp diff manually. User creation is the slowest operation - adding
groups/memberships is much faster, so we don't need to log as
frequently.

Note that there is a usability trade-off on how frequently we log
depending on whether the user is using the slower (but more common)
method of going via LDAP, vs the much faster (but more obscure) method
of writing directly to sam.ldb with ldb:nosync=true. In my tests, we end
up logging every ~30-ish secs with LDAP, and every ~3 seconds with
direct file writes.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
71c66419bb traffic_replay: Convert print() to logger.info()
Using logger is more helpful here because it includes timestamps, so we
can see how long things are taking. It's also more consistent with the
rest of the traffic_replay logging.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
32e58227cd traffic_replay: Write group memberships once per group
Each user-group membership was being written to the DB in a single
operation. With large numbers of users (e.g. 10,000 in average 15 groups
each), this becomes a lot of operations (e.g. 150,000). This patch
reworks the code so that we write the memberships for a group in
one operation. E.g. instead of 150,000 DB operations, we might make
1,500. This makes writing the group memberships several times
faster.

Note that rthere is a performance vs memory tradeoff. When we hit
10,000+ members in a group, memory-usage in the underlying DB modify
operation becomes very inefficient/costly. So we avoid potential memory
usage problems by writing no more than 1,000 users to a group at once.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
a29ee3a745 traffic_replay: Re-organize assignments to be group-based
We can speed up writing the group memberships by adding multiple users
to a group in a single DB modify operation.

To do this, we first need to reorganize the assignments so instead
of being a set of tuples, it's a dictionary where key=group and
value=list-of-users-in-group.

add_users_to_groups() now iterates through the users/groups slightly
differently, but mostly it's just indentation changes. We haven't
changed the number of DB operations yet - we'll do that in the next
patch.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
5ad7fc7335 traffic_replay: Prevent users having 1000+ memberOf links
When adding 10,000 users, one user would end up in over 1000 groups.
With 100,000 users, it would be more like 10,000 groups. While it makes
sense to have groups with large numbers of users, having a single user
in 1000s of groups is probably less realistic.

This patch changes the shape of the Pareto distribution that we use to
assign users to groups. The aim is to cap users at belonging to at most
~500 groups. Increasing the shape of the Pareto distribution pushes the
user assignments so they're closer to the average, and the tail (with
users in lots of groups) is not so large).

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
fdd75407af traffic_replay: Change user distribution to use Pareto Distribution
The current probability we were assigning to users roughly approximates
the Pareto Distribution (with shape=1.0). This means the code now uses a
documented algorithm (i.e. explanation on Wikipedia). It also allows us
to vary the distribution by changing the shape parameter.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
898e6b4332 traffic_replay: Improve assign_groups() performance with large domains
When assigning 10,000 users to 15 groups each (on average),
assign_groups() would take over 30 seconds. This did not include any DB
operations whatsoever. This patch improves things, so that it takes less
than a second in the same situation.

The problem was the code was looping ~23 million times where the
'random.random() < probability * 10000' condition was not met. The
problem is individual group/user probabilities get lower as the number
of groups/users increases. And so with large numbers of users, most of
the time the calculated probability was very small and didn't meet the
threshold.

This patch changes it so we can select a user/group in one go, avoiding
the need to loop multiple times.

Basically we distribute the users (or groups) between 0.0 and 1.0, so
that each user has their own 'slice', and this slice is proporational to
their weighted probability. random.random() generates a value between
0.0 and 1.0, so we can use this to pick a 'slice' (or rather, we use
this as an index into the list, using .bisect()). Users/groups with
larger probabilities end up with larger slices, so are more likely to
get picked.

The end result is roughly the same distribution as before, although the
first 10 or so user/groups seem to get picked more frequently, so the
weighted-probability calculations may need tweaking some more.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
18740ec0dd traffic_replay: Split out random group membership generation logic
This doesn't change functionality at all. It just moves the probability
calculations out into separate functions.

We want to tweak the logic/implementation behind this code, but the
rest of assign_groups() doesn't really care how the underlying
probabilities are worked out, so long as it gets a suitably random
user/group membership each time round the loop.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:16 +01:00
Tim Beale
e3e84b0f6d traffic_replay: Add helper class for group assignments
Wrap up the group assignment calculations in a helper class. We're going
to tweak the internals a bit in subsequent patches, but the rest of the
code doesn't really need to know about these changes.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-11-04 23:55:15 +01:00
Tim Beale
4943473102 traffic_replay: Change print() to use logger()
This reduces noise, so the messages only come out if you specify
--debug.

Signed-off-by: Tim Beale <timbeale@catalyst.net.nz>
Reviewed-by: Douglas Bagnall <douglas.bagnall@catalyst.net.nz>
2018-10-31 00:30:16 +01:00