[LLVMdev] Proposal: change LNT’s regression detection algorithm and how it is used to reduce false positives

[Kristof Beyls]

Thanks for raising this, Chris!

 

I also think that improving the signal-to-noise ratio in the performance
reports produced by LNT are essential to make the performance-tracking
bots useful and effective.

 

Our experience, using LNT internally, has been that if the number of false
positives are low enough (lower than about half a dozen per report or day),
they become useable, leaving only a little bit of manual investigation work
to detect if a particular change was significant or in the noise. Yes, ideally
the automated noise detection should be perfect; but even if it's not perfect,
it will already be a massive win.

 

I have some further ideas and remarks below.

 

Thanks,

 

Kristof

 

 

> -----Original Message-----

> From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu]

> On Behalf Of Chris Matthews

> Sent: 15 May 2015 22:25

> To: LLVM Developers Mailing List

> Subject: [LLVMdev] Proposal: change LNT’s regression detection algorithm

> and how it is used to reduce false positives

> 

> tl;dr in low data situations we don’t look at past information, and that

> increases the false positive regression rate.  We should look at the

> possibly incorrect recent past runs to fix that.

> 

> Motivation: LNT’s current regression detection system has false positive

> rate that is too high to make it useful.  With test suites as large as

> the llvm “test-suite” a single report will show hundreds of regressions.

> The false positive rate is so high the reports are ignored because it is

> impossible for a human to triage them, large performance problems are

> lost in the noise, small important regressions never even have a chance.

> Later today I am going to commit a new unit test to LNT with 40 of my

> favorite regression patterns.  It has gems such as flat but noisy line,

> 5% regression in 5% noise, bimodal, and a slow increase, we fail to

> classify most of these correctly right now. They are not trick

> questions, all are obvious regressions or non-regressions, that are

> plainly visible. I want us to correctly classify them all!

 

That's a great idea!

Out of all of the ideas in this email, I think this is the most important
one to implement first.

 

> Some context: LNTs regression detection algorithm as I understand it:

> 

> detect(current run’s samples, last runs samples) —> improve, regress or

> unchanged.

> 

>     # when recovering from errors performance should not be counted

>     Current or last run failed -> unchanged

> 

>     delta = min(current samples) - min(prev samples)

 

I am not convinced that "min" is the best way to define the delta.

It makes the assumption that the "true" performance of code generated by llvm
is the fastest it was ever seen running. I think this isn't the correct way
to model e.g. programs with bimodal behaviour, nor programs with a normal
distribution. I'm afraid I don't have a better solution, but I think the
Mann Whitney U test - which tries to determine if the sample points seem
to indicate different underlying distributions - is closer to what we
really ought to use to detect if a regression is "real". This way, it models

that a fixed program, when run multiple times, has a distribution of

performance. I think that using "min" makes too many broken assumptions on

what the distribution can look like.

 

> 

>     # too small to measure

>     delta <  (confidence*machine noise threshold (0.0005s by default)) -

> > unchanged

> 

>     # too small to care

>     delta % < 1% -> unchanged

> 

>     # too small to care

>     delta < 0.01s -> unchanged

> 

>     if len(current samples) >= 4 && len(prev samples) >= 4

>          Mann whitney U test -> possible unchanged

> 

>     #multisample, confidence interval check

>     if len(current samples) > 1

>            check delta within samples confidence interval -> if so,

> unchanged, else Improve, regress.

> 

>     # single sample,range check

>     if len(current samples) == 1

>         all % deltas above 1% improve or regress

> 

> 

> The too small to care rules are newer inventions.

> 

> Effectiveness data: to see how well these rules work I ran a 14 machine,

> 7 day report:

> 

> - 16773 run comparisons

> - 13852 marked unchanged because of small % delta

> - 2603 unchanged because of small delta

> - 0 unchanged because of Mann Whitney U test

> - 0 unchanged because of confidence interval

> - 318 improved or regressed because single sample change over 1%

> 

> Real regressions: probably 1 or 2, not that I will click 318 links to

> check for sure… hence the motivation.

> 

> Observations: Most of the work is done by dropping small deltas.

> Confidence intervals and Mann Whitney U tests are the tests we want to

> be triggering, however they only work with many samples. Even with

> reruns, most tests end up being a single sample.  LNT bots that a

> triggered after another build (unless using the multisample feature)

> just have one sample at each rev.  Multisample is not a good option

> because most runs already take a long time.

> 

> Even with a small amount of predictable noise, if len(current samples)

> == 1, will flag a lot of samples, especially if len(prev) > 1.  Reruns

> actually make this worse by making it likely that we flag the next run

> after the run we rerun.  For instance, a flat line with 5% random noise

> flags all the time.

> 

> Besides the Mann Whitney U test, we are not using prev_samples in any

> way sane way.

> 

> Ideas:

> 

> -Try and get more samples in as many places as possible.  Maybe —

> multisample=4 should be the default?  Make bots run more often (I have

> already done this on green dragon).

 

FWIW, the Cortex-A53 performance tracker I've set up recently uses
multisample=3. The Cortex-A53 is a slower/more energy-efficient core,
so it takes about 6 hours to do a LLVM rebuild + 3 runs of the LNT
benchmarks (see http://llvm.org/perf/db_default/v4/nts/machine/39).

 

> - Use recent past run information to enhance single sample regression

> detection.  I think we should add a lookback window, and model the

> recent past.  I tired a technique suggested by Mikhail Zolotukhin of

> computing delta as the smallest difference between current and all the

> previous samples.  It was far more effective.  Alternately we could try

> a confidence interval generated from previous, though that may not work

> on bimodal tests.

 

The noise levels per individual program are often dependent on the

micro-architecture of the core it runs on. Before setting up the Cortex-A53

performance tracking bot, I've done a bit of analysis to find out what the noise

levels are per program across a Cortex-A53, a Cortex-A57 and a Core i7 CPU. Below

is an example of a chart for just one program, indicating that the noise level is
sometimes dependent on the micro-architecture of the core it runs on. Whereas a

Mann-Withney U - or similar - test would probably find - given enough data

points - what should be considered noise and what not; there may be a way to

run the test-suite in benchmark mode many times when a board gets set up, and analyse

the results of that. The idea is that this way, the noisiness of the board as setup

for fixed binaries could be measured, and that information could be used when not

enough sample points are available.

(FWIW: for this program, the noisiness seems to come from noisiness in the number
of branch mispredicts).

BTW – graphs like the one below make me think that the LNT webUI should be showing
sample points be default instead of line graphs showing the minimum execution time
per build number.

 



 

> - Currently prev_samples is almost always just one other run, probably

> with only one sample itself.  Lets give this more samples to work with.

> Start passing more previous run data to all uses of the algorithm, in

> most places we intentionally limit the computation to current=run and

> previous=run-1, lets do something like previous=run-[1-10]. The risk in

> this approach is that regression noise in the look back window could

> trigger a false negative (we miss detecting a regression).  I think this

> is acceptable since we already miss lots of them because the reports are

> not actionable.

> 

> - Given the choice between false positive and false negative, lets err

> towards false negative.  We need to have manageable number of

> regressions detected or else we can’t act on them.

 

This sounds like a good idea to me. Let's first make sure we have a working
system of (semi-?)automatically detecting at least a good portion of the
significant performance regression. After that we can fine tune to reduce
false negatives to catch a larger part of all significant performance
regressions.

 

> 

> Any objections to me implementing these ideas?

 

Absolutely not. Once implemented, we probably ought to have an idea about how
to test which combination of methods works best in practice. Could the
sample points you’re going to add to the LNT unit tests help in testing which
combination of methods work best?

 

I've got 2 further ideas, based on observations from the data coming from the
Cortex-A53 performance tracker that I added about 10 days ago - see
http://llvm.org/perf/db_default/v4/nts/machine/39.

I'll be posting patches for review for these soon:

 

1. About 20 of the 300-ish programs that get run in benchmark-only mode run
for less than 10 milliseconds. These 20 programs are one of the main sources
of noisiness. We should just not run these programs in benchmark-only mode.

Or - alternatively we should make them run a bit longer, so that they are less
noisy.

 

2. The board I'm running the Cortex-A53 performance tracker on is a big.LITTLE
system with 2 Cortex-A57s and 4 Cortex-A53s. To build the benchmark binaries,
I'm using all cores, to make the turn-around time of the bot as fast as possible.
However, this leads to huge noise levels on the "compile_time" metric, as sometimes
a binary gets compiled on a Cortex-A53 and sometimes on a Cortex-A57. For this
board specifically, it just shouldn't be reporting compile_time at all, since the
numbers are meaningless from a performance-tracking use case.

 

 

Another thought: if we could reduce the overall run-time of the LNT run in
benchmark-only mode, we could run more "multi-samples" in the same amount of
time. I did a quick analysis on whether it would be worthwhile to invest effort
in making some of the long-running programs in the test-suite run shorter in
benchmarking mode. On the Cortex-A53 board, it shows that the 27 longest-running
programs out of the 300-ish consume about half the run-time. If we could easily
make these 27 programs run an order-of-magnitude less long, we could almost halve
the total execution time of the test-suite, and hence run twice the number of
samples in the same amount of time. The longest running programs I’ve found are,
sorted:

 

  0. 7.23% cumulative (7.23% - 417.15s this program) nts.MultiSource/Benchmarks/PAQ8p/paq8p.exec

  1. 13.74% cumulative (6.51% - 375.84s this program) nts.MultiSource/Benchmarks/SciMark2-C/scimark2.exec

  2. 18.83% cumulative (5.08% - 293.16s this program) nts.SingleSource/Benchmarks/Polybench/linear-algebra/kernels/symm/symm.exec

  3. 21.60% cumulative (2.77% - 160.02s this program) nts.MultiSource/Benchmarks/mafft/pairlocalalign.exec

  4. 24.01% cumulative (2.41% - 138.98s this program) nts.SingleSource/Benchmarks/CoyoteBench/almabench.exec

  5. 26.32% cumulative (2.32% - 133.59s this program) nts.MultiSource/Applications/lua/lua.exec

  6. 28.26% cumulative (1.94% - 111.80s this program) nts.MultiSource/Benchmarks/ASC_Sequoia/IRSmk/IRSmk.exec

  7. 30.11% cumulative (1.85% - 106.56s this program) nts.MultiSource/Benchmarks/ASC_Sequoia/AMGmk/AMGmk.exec

  8. 31.60% cumulative (1.49% - 86.00s this program) nts.SingleSource/Benchmarks/CoyoteBench/huffbench.exec

  9. 32.75% cumulative (1.15% - 66.37s this program) nts.MultiSource/Benchmarks/TSVC/NodeSplitting-dbl/NodeSplitting-dbl.exec

10. 33.90% cumulative (1.15% - 66.13s this program) nts.MultiSource/Applications/hexxagon/hexxagon.exec

11. 35.04% cumulative (1.14% - 65.98s this program) nts.SingleSource/Benchmarks/Polybench/linear-algebra/kernels/syr2k/syr2k.exec

12. 36.14% cumulative (1.10% - 63.21s this program) nts.MultiSource/Benchmarks/TSVC/IndirectAddressing-dbl/IndirectAddressing-dbl.exec

13. 37.22% cumulative (1.08% - 62.35s this program) nts.SingleSource/Benchmarks/SmallPT/smallpt.exec

14. 38.30% cumulative (1.08% - 62.30s this program) nts.MultiSource/Benchmarks/nbench/nbench.exec

15. 39.37% cumulative (1.07% - 61.98s this program) nts.MultiSource/Benchmarks/TSVC/ControlFlow-dbl/ControlFlow-dbl.exec

16. 40.40% cumulative (1.03% - 59.50s this program) nts.MultiSource/Applications/SPASS/SPASS.exec

17. 41.37% cumulative (0.97% - 55.74s this program) nts.MultiSource/Benchmarks/TSVC/Expansion-dbl/Expansion-dbl.exec

18. 42.33% cumulative (0.96% - 55.40s this program) nts.SingleSource/Benchmarks/Misc/ReedSolomon.exec

19. 43.27% cumulative (0.94% - 54.34s this program) nts.MultiSource/Benchmarks/TSVC/IndirectAddressing-flt/IndirectAddressing-flt.exec

20. 44.21% cumulative (0.94% - 54.20s this program) nts.MultiSource/Benchmarks/TSVC/StatementReordering-dbl/StatementReordering-dbl.exec

21. 45.12% cumulative (0.91% - 52.46s this program) nts.SingleSource/Benchmarks/Polybench/datamining/covariance/covariance.exec

22. 46.01% cumulative (0.89% - 51.49s this program) nts.MultiSource/Benchmarks/ASC_Sequoia/CrystalMk/CrystalMk.exec

23. 46.89% cumulative (0.88% - 50.66s this program) nts.MultiSource/Benchmarks/TSVC/ControlFlow-flt/ControlFlow-flt.exec

24. 47.73% cumulative (0.84% - 48.74s this program) nts.MultiSource/Benchmarks/TSVC/CrossingThresholds-dbl/CrossingThresholds-dbl.exec

25. 48.57% cumulative (0.84% - 48.43s this program) nts.MultiSource/Benchmarks/TSVC/InductionVariable-dbl/InductionVariable-dbl.exec

26. 49.40% cumulative (0.83% - 47.92s this program) nts.SingleSource/Benchmarks/Polybench/datamining/correlation/correlation.exec

27. 50.22% cumulative (0.81% - 46.92s this program) nts.MultiSource/Benchmarks/TSVC/NodeSplitting-flt/NodeSplitting-flt.exec

28. 51.03% cumulative (0.81% - 46.90s this program) nts.MultiSource/Applications/minisat/minisat.exec

29. 51.81% cumulative (0.78% - 44.88s this program) nts.MultiSource/Benchmarks/TSVC/Packing-dbl/Packing-dbl.exec

…

 

 

For example, there seem to be a lot of TSVC benchmarks in the longest running ones.
They all seem to take a command line parameter to define the number of iterations the main
loop in the benchmark should be running. Just tuning these, so all these benchmarks runs
O(1s) would make the overall test-suite already run significantly faster.

 

For the Polybench test cases: they print out lots of floating point numbers – this
probably should be changed in the makefile so they don’t dump the matrices they work
on anymore. I’m not sure how big the impact will be on overall run time for the Polybench
benchmarks when doing this.

 

 

 

 

 

 

 

 

> _______________________________________________

> LLVM Developers mailing list

>  <mailto:LLVMdev at cs.uiuc.edu> LLVMdev at cs.uiuc.edu          <http://llvm.cs.uiuc.edu> http://llvm.cs.uiuc.edu

>  <http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150518/21963a0d/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image003.png
Type: image/png
Size: 46772 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150518/21963a0d/attachment.png>