3:15 a.m.
Hi William,
Thanks for these very interesting results.
It would help if you can provide the stap scripts to make sure what you
are accounting the latency.
Also just to be sure is latency a synonym for response time ?
Regarding the comparison (tail) 1client/16client. It looks to me that
the tail are equivalent: The more we have clients the more we have long
latency. So in a first approach I would eliminate contention effect.
Regarding the ratio shorter/longer latency (assuming the search are
equivalent) this is interesting to know why we have such effect. One of
the possible cause I was thinking of is the impact of DB thread
(checkpointing or trickling). But if it exists similar long tail in
ldbm_back, the absolute value is much lower than the opshare_search: in
short ldbm_back accounted at most for 4ms while opshared for 67ms. So
there is something else (aci, network, frontend..).
Regarding USDT I think it is very good idea. However, just to share some
recent stap experience, I found it intrusive. In short, I had not the
same throughput with and without. In my case it was not a problem, as I
wanted to investigate a reproducible contention. But if we want
support/user/customers to use it, the performance impact in production
will be valid point.
best regards
thierry
On 12/9/19 6:16 AM, William Brown wrote:
Hi all,
Following last weeks flamegraph runs, I wanted to find more details on exactly what was
happening. While flamegraphs highlighted that a changed existed between single and
multithreaded servers, it did not help to isolate where
the change was occuring.
To understand this I have started to work on a set of systemtap scripts that we can use
to profile 389ds. These will be included in a future PR.
Here are the hisograms from an initial test of profiling "do_search"
1 thread
stap test_do_search.stp
^CDistribution of latencies (in nanoseconds) for 441148 samples
max/avg/min: 35911542/85694/38927
value |-------------------------------------------------- count
8192 | 0
16384 | 0
32768 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 167285
65536 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 239280
131072 |@@@@@ 25788
262144 |@ 8530
524288 | 252
1048576 | 6
2097152 | 1
4194304 | 3
8388608 | 0
16777216 | 2
33554432 | 1
67108864 | 0
134217728 | 0
16 thread
stap test_do_search.stp
^CDistribution of latencies (in nanoseconds) for 407806 samples
max/avg/min: 100315928/112407/39368
value |-------------------------------------------------- count
8192 | 0
16384 | 0
32768 |@@@@@@@@@@@@@@@@@@@@ 100281
65536 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 249656
131072 |@@@@@@@ 37837
262144 |@@@ 18322
524288 | 1171
1048576 | 203
2097152 | 90
4194304 | 74
8388608 | 83
16777216 | 58
33554432 | 21
67108864 | 10
134217728 | 0
268435456 | 0
It's interesting to note the tail latency here: On the 16 thread version we see 67000
less in the 32768 buckets, shifting mostly through the 131072 and 262144 buckets, as well
as showing a much greater number of calls in the tail. In thread 1 no operation made it to
67108864, but 10 did in 16thread, along with ~200 more that are higher than 1048567, and
~1500 more that are greater than 524288. This kind of tailing means we have
"spikes" of latency throughout the execution, which then have a minor flow on
cause other operations to be increased in latency.
These are all in nanoseconds, so this means most operations are in do_search for
0.000131072 seconds or less, while there are spikes of operations taking between nearly
0.001048576 and 0.067108864 seconds to complete (which is an 8x to 512x increase in
execution time.
From these point I took two measurements of back_ldbm and the access log, knowing these
were easy targets for potential areas of latency. Both of these were performed with the
16thread server. I didn't need to do this on the 1 thread because I'm looking for
tail latency, not comparisons now. Because we know that there is an increase of ~1500
events of high latency, we are looking for similar numbers to appear in other tail
latencies.
^CDistribution of back_ldbm latencies (in nanoseconds) for 364223 samples
max/avg/min: 4757744/31192/13622
value |-------------------------------------------------- count
2048 | 0
4096 | 0
8192 |@@ 14318
16384 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 261265
32768 |@@@@@@@@@@@@ 63533
65536 |@@@ 20624
131072 | 4062
262144 | 304
524288 | 68
1048576 | 33
2097152 | 15
4194304 | 1
8388608 | 0
16777216 | 0
This shows some tail latency in back_ldbm however it's infrequent, only showing a
handful of spikes, but they do exist and tend to taper off quickly as they approach
2097152. Certainly these should be looked at as they will likely add up to affecting
do_search.
stap -v test_access_log.stp
^CDistribution of slapi_log_access latencies (in nanoseconds) for 1284350 samples
max/avg/min: 23843992/5138/1036
value |-------------------------------------------------- count
256 | 0
512 | 0
1024 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 379710
2048 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 361243
4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 412480
8192 |@@@@@@@@@@ 83815
16384 |@@@@ 40087
32768 | 5076
65536 | 1253
131072 | 406
262144 | 130
524288 | 72
1048576 | 33
2097152 | 16
4194304 | 19
8388608 | 1
16777216 | 9
33554432 | 0
67108864 | 0
From this, I can see that while some tail latency exists in some conditions of
back_ldbm, there is a much longer and present tail in the access log, with nearly 2000
operations in or exceeding the 65536 mark - remember, in do_search most of the operations
as a whole take 65536, so we have latency spikes in slapi_log_access that are as large or
larger than whole search operations as a total.
Now we can see "most! operations are fimly below 16384. This is what we want to see,
but it's interesting that a slapi_log_access can "take up to" a quarter of a
whole operations processing on the avg case.
What's truly telling here is the tail down to 16777216, with ~250 operations
exceeding 262144 ns. This shows that within slapi_log_access, we have have large spiked
delays in the behaviour, which will be affecting both the tail latency of do_search as a
whole, but also causing other slapi_log_access operations to "hold up".
My next steps from here will be:
* To add USDT probes to the logs and back_ldbm to get better, fine detail about what is
causing the excessive latency.
* these probes are also needed to resolve what appears to be a symbol resolution issue
with systemtap when optimisations are enabled.
* To add probes in other parts of the code base to get better visualisation about where
and how long timings are taking through an operation.
* To run a lock contention profile (I was unable to do this today due to bugs in
systemtap)
* To document the setup proceedures
* Commit all these into a PR
* Document some actionable improvements we can make to improve the server performance.
Hope that is interesting to everyone,
--
Sincerely,
William
_______________________________________________
389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/
List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines
List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...