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

[chris_matthews]

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!

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)

    # 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).

- 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.

- 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.

Any objections to me implementing these ideas?

[_sean_silva]

Is there a way to download the data off http://llvm.org/perf? I’d like to help with this but I don’t have a good dataset to analyze.

It definitely seems like the weakest part of the current and proposed scheme is that it only looks at two runs. That is basically useless when we’re talking about only a handful of samples (<4???) per run. Since the machine’s noise can be modeled from run to run (also sample to sample, but for simplicity just consider run to run) as a random process in the run number, all the techniques from digital filtering come into play. From looking at a couple of the graphs on LNT, the machine noise appears to be almost exclusively at Nyquist (i.e. it alternates from sample to sample) falling down to a bit at half Nyquist (I can analyze in more detail if I can get my hands on the data). We probably want a lowpass differentiator at about half Nyquist.

I would strongly recommend starting with a single benchmark on a single machine and coming up with detection routine just for it that is basically 100% accurate, then generalizing as appropriate so that you are getting reliable coverage of a larger portion of the benchmarks. The machine’s noise is probably easiest to characterize and most generalizable across runs.

– Sean Silva

[chris_matthews]

Easiest way to get the data off llvm.org/perf is to use the json APIs. For many of the pages if you pass a &json=True LNT will give you a json reply. For example, go to a run page, a click the check boxes next to a bunch of runs and click graph. When all the run lines popup in the graph page, add json=True and it will download data for those tests on those machines.

For example, on the O3 tester, all the benchmarks in Multisource/Applications:

http://llvm.org/perf/db_default/v4/nts/graph?plot.1327=21.1327.0&plot.1053=21.1053.0&plot.1232=21.1232.0&plot.1483=21.1483.0&plot.1014=21.1014.0&plot.1138=21.1138.0&plot.1180=21.1180.0&plot.1288=21.1288.0&plot.1129=21.1129.0&plot.1425=21.1425.0&plot.1456=21.1456.0&plot.1038=21.1038.0&plot.1452=21.1452.0&plot.1166=21.1166.0&plot.1243=21.1243.0&plot.1116=21.1116.0&plot.1326=21.1326.0&plot.1279=21.1279.0&plot.1007=21.1007.0&plot.1394=21.1394.0&plot.1017=21.1017.0&plot.1443=21.1443.0&plot.1445=21.1445.0&plot.1197=21.1197.0&plot.1332=21.1332.0&json=True

I have scripts for converting the json data into Python Pandas format, though the format is so simple you can really parse it with anything. I could contribute them if anyone would find them helpful. I know there are also scripts floating around for scraping all machines and runs from the LNT instance by first looking up the run and machine list, then fetching all the tests for each.

[chris_matthews]

And maybe just while I’m poking around, some runs that I see flagged regularly:

http://llvm.org/perf/db_default/v4/nts/graph?plot.0=21.1327.0&highlight_run=26921
http://llvm.org/perf/db_default/v4/nts/graph?plot.0=21.1432.0&highlight_run=26921
http://llvm.org/perf/db_default/v4/nts/graph?plot.0=21.1222.0&highlight_run=26921
http://llvm.org/perf/db_default/v4/nts/graph?plot.0=21.1122.1&highlight_run=26921

[_sean_silva]

Thanks for the info!

I actually scraped by (no pun intended) for a first look by scraping the source of http://llvm.org/perf/db_default/v4/nts/graph?hide_lineplot=yes&show_points=yes&moving_window_size=10&plot.0=21.1419.0&submit=Update#
which I happened to choose. I’ll definitely take advantage of the JSON API in the future!

Do you know how the individual data points correspond to the actual samples? With “show all sample points” multiple points are shows for at each revision tested, suggesting that the data is somewhere. Also, “show all sample points” does not show the multiplicity at all, so it is hard to tell what each point is.

If you’re interested, the benchmark I looked at appears to have a similar behavior to some of the ones you mentioned downthread, so you might find it interesting:
https://drive.google.com/file/d/0B8v10qJ6EXRxdWFIN1dnYWVQWGc/view?usp=sharing
A lowpass differentiator with a cutoff of about Pi/10 seems to do a good job for that machine’s noise characteristics. The big bimodal jumps of course need their own special cleanup before applying it though.
Also, it appears that the noise spectrum changes in ways that are not correlated with the major performance changes… do you know if that machine has had its configuration changed?
It would be useful to track some other data points along with just the time, like total cycles, total insts retired, total icache misses, branch prediction ratio, page faults, and maybe a few other things just so that we have a few more degrees of freedom along which to look for verification.

– Sean Silva

[kbeyls]

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

From: llvmdev-bounces@cs.uiuc.edu [mailto:llvmdev-bounces@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.

[kbeyls]

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

From: llvmdev-bounces@cs.uiuc.edu [mailto:llvmdev-bounces@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.

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).
BTW, what is "green dragon"?

- 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. In
attachment 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 in attachment make me think that the LNT webUI should be showing
sample points by 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.

[Michael_Zolotukhin]

Hi Chris and others!

I totally support any work in this direction.

In the current state LNT’s regression detection system is too noisy, which makes it almost impossible to use in some cases. If after each run a developer gets a dozen of ‘regressions’, none of which happens to be real, he/she won’t care about such reports after a while. We clearly need to filter out as much noise as we can - and as it turns out even simplest techniques could help here. For example, the technique I used (which you mentioned earlier) takes ~15 lines of code to implement and filters out almost all noise in our internal data-sets. It’d be really cool to have something more scientifically-proven though:)

One thing to add from me - I think we should try to do our best in assumption that we don’t have enough samples. Of course, the more data we have - the better, but in many cases we can’t (or we don’t want) to increase number os samples, since it dramatically increases testing time. That’s not to discourage anyone from increasing number of samples, or adding techniques relying on a significant number of samples, but rather to try mining as many ‘samples’ as possible from the data we have - e.g. I absolutely agree with your idea to pass more than 1 previous run.

Thanks,
Michael

[_sean_silva]

Hi Chris and others!

I totally support any work in this direction.

In the current state LNT’s regression detection system is too noisy, which
makes it almost impossible to use in some cases. If after each run a
developer gets a dozen of ‘regressions’, none of which happens to be real,
he/she won’t care about such reports after a while. We clearly need to
filter out as much noise as we can - and as it turns out even simplest
techniques could help here. For example, the technique I used (which you
mentioned earlier) takes ~15 lines of code to implement and filters out
almost all noise in our internal data-sets. It’d be really cool to have
something more scientifically-proven though:)

One thing to add from me - I think we should try to do our best in
assumption that we don’t have enough samples. Of course, the more data we
have - the better, but in many cases we can’t (or we don’t want) to
increase number os samples, since it dramatically increases testing time.

Why not just start out with only a few samples, then collect more for
benchmarks that appear to have changed?

-- Sean Silva

[chris_matthews]

The reruns flag already does that. It helps a bit, but only as long as the the benchmark is flagged as regressed.

[chris_matthews]

In r237661 I committed an initial set of tests. All the failing tests are commented out right now (python 2.7 does not have xfail). I extracted the data sets to the top of the file so they are easy to paste into ipython etc.

[chris_matthews]

I agree. Fixed in r237748.

[kbeyls]

Thanks!

I guess someone will need to restart the LNT server at llvm.org/perf to pick up this change?

Renato, do you have access to restart the LNT server at llvm.org/perf?

Kristof

[Renato_Golin1]

I do, but there are other services running on that machine. I'd prefer
if Tanya or someone else would restart it.

cheers,
--renato

[_sean_silva]

I found an interesting datapoint:

In the last 10,000 revisions of LLVM+Clang, only 10 revisions actually caused the binary of MultiSource/Benchmarks/BitBench/five11 to change. So if just store a hash of the binary in the database, we should be able to pool all samples we have collected while the binary is the the same as it currently is, which will let us use significantly more datapoints for the reference.

Also, we can trivially eliminate running the regression detection algorithm if the binary hasn’t changed.

– Sean Silva

[Smith_Kevin_B]

Ø Also, we can trivially eliminate running the regression detection algorithm if the binary hasn’t changed.

Strongly agree. Internal performance regression testing for Intel compiler uses this method to help eliminate noise. It is

a great first line method to greatly cut down on developer time wasted chasing phantoms.

Kevin Smith

[kbeyls]

Sounds like a great idea!

[Renato_Golin1]

In the last 10,000 revisions of LLVM+Clang, only 10 revisions actually
caused the binary of MultiSource/Benchmarks/BitBench/five11 to change. So if
just store a hash of the binary in the database, we should be able to pool
all samples we have collected while the binary is the the same as it
currently is, which will let us use significantly more datapoints for the
reference.

+1

Also, we can trivially eliminate running the regression detection algorithm
if the binary hasn't changed.

+2!

--renato

[chris_matthews]

I agree this is a great idea. I think it needs to be fleshed out a little though.

It would still be wise to run the regression detection algorithm, because the test suite changes and the machines change, and the algorithm is not perfect yet. It would be a valuable source of information though.

This is not a small change to how LNT works, so I think some due diligence is necessary. Is clang *really* that deterministic, especially over successive revs? I know it is supposed to be. Does anyone have any data to show this is going to be an effective approach? It seems like there are benchmarks in the test-suite which use __DATE__ and __TIME__ in them. I assume that will be a problem?

[Renato_Golin1]

We also need to consider multiple binaries, like shared libraries that are compiled, for instance, libc++.

Maybe a simple binary diff + ldd + binary diff on all deps would work…

Cheers,
Renato