Re: Feedback request regarding HSOC project (bringing sanity to the GHC performance test-suite)

Hi Jared, On Tue, Sep 5, 2017, 19:39 Jared Weakly wrote:

Hi Tamar,

That framework failure is due to a somewhat embarrassing error that I thought I had caught earlier; line 298 shouldn't have existed (it was a small mistake from converting the all.T file from using the old function to using the new collect_stats function. I have fixed this and it will be pushed by the time you read this email. That being said, the individual tests or units are very isolated and a framework failure simply means that unit didn't get run; in this case it means that entire all.T file didn't get run since the error was in reading the file, but the rest of the files should've had their performance tests recorded properly.

The .git/refs/notes/perf is an implementation detail. Git notes have the concept of namespaces; so, in order to avoid cluttering up a global namespace in git notes with stuff only the performance tests will use, all the performance metrics are stored in the namespace 'perf'.

This format of the git notes is mentioned in the code for the testsuite but I will make this more visible in the README and other documentation. The format of the git notes is: $test-env $test_name $test_way $metric_measured $value_collected (separated by tabs)

The maximum deviation the test allows is inside the respective all.T's; additionally, if you set the verbosity level of the test-suite to a value >= 4, you will see the expected value, allowed deviation, lower bound, upper bound, the actual measured value, and (if the test fell outside the bounds) how much the actual value deviated from the expected value. This information will also print if the test falls outside of the allowed bounds.

perf_notes.py exists as both an internal library and a measurement tool (hopefully to be useful to developers). You can give the tool several commits and it will give you a comparison of the union of all the tests in those commits, with an output very similar in style to noFib. I imagine this will be useful mostly to people who want to improve the performance of the compiler so they can see which tests have regressed the most over time (or which have improved the most over time); but as it works over commits, it can also be useful for a developer wanting to know if they've made a measurable difference with their patch.

The notes are updated every time the testsuite is ran. However they are updated only at the very end of the execution of the testsuite in a single command (the information is collected in a python datastructure which is turned into a string and given to git notes). This behavior means that if the testsuite is ran more than once in-between commits that 'duplicate values' will exist in the git notes. I'm not quite sure how to deal with this yet; I am considering just grabbing the latest value if multiples exist. This also means you can test just one test and then run other tests and have those values added into the git notes without losing your older values which is why the behavior is kept this way. (I will make sure this is more prominent in the docs somewhere).

The note update is done using python's subprocess library. I have no idea how resilient that is to git failure; I'd imagine that if it was busy it would just silently fail to update. Fortunately, the update process is as close to atomic as one can get. I'll see if I can figure out a way to force a repo lock to test this out. I'm open to suggestions as to how to deal with this better and I'll also google around and see if anyone has a good solution.

I think you can just wait and retry and do so a specified max number of r times. That should be good enough for most cases. Unless the process on the other end has died but then the user needs to clean up the lock firsts.

Platform discrepancies are completely sidestepped because of the way git notes work. The performance metrics are entirely local and stay on your computer; they won't be pushed or shared with any other users. That means that the performance numbers are completely tailored to your platform so there is effectively an 0% margin of "OS-related" error that needs to be accounted for. The collect_stats function is very much designed to be declarative and "set it and forget it". As such, the need to even record values at all is obsoleted (one of the main motivators of this project in the first place).

This I don't quite understand. If I get this right. It means I will now always have to run the full performance benchmark suite for each change I have twice? Before and after locally? I had thought the notes would be pushed as I would find it useful to have the perfect history locally if I wanted to. Performance by its very nature is very platform specific, I feel that this change makes it harder for the platforms we don't have a CI for to run benchmarks. So basically only Linux. This would be unfortunate as it would mean we would effectively stop tracking performance on e.g. Windows and Mac OS since the current implementation doesn't allow for the data to live together in the same repo.

Hopefully this answers some questions; I'll make sure this sort of information is available somewhere so that later users can find these answers again. Thanks for your thoughts! They were very helpful.

Regards, Jared

...

Hi Jared,

First off, thanks for all the hard work on this. I checked out your branch and made a run, I noticed at the end it had

Framework failures: . ./perf/compiler/all.T [] (unexpected indent (<string>, line 298))

so I assume none of the perf tests were run?

Though I do see a .git/refs/notes/perf, so I assume your ref is is perf?

Doing a git notes --ref perf show I see somethings were collected at some point

local T3924 normal bytes allocated 47064 local haddock.base normal bytes allocated 18427047160 local haddock.Cabal normal bytes allocated 15863910848 local haddock.compiler normal bytes allocated 50656428952

which brings me up to my first question, I'm guessing the number here is

...

number of bytes allocated for the test? Is there a way for me to see what the maximum deviation the test allows is and how far I am from it? Do I just get the information like before only when a test fails? How does

...

look like? Same as before?

It's also not entirely clear to be what perf_notes.py can be used, is it just an infrastructure tool? or is it something you foresee as useful for a developer?

lastly, how often do you update notes? It's probably too late for this now, but git, especially msys git can be especially slow, so I would have

On Mon, Sep 4, 2017 at 10:02 PM, Phyx wrote: the that liked

...

the notes to be updated in batches to not slow down the testsuite run on Windows.

Which brings me to my next question, how resilient are you to failures updating git? some IDE/environments like vscode automatically issue git operations in the background. so git may be busy when you try to update and the operation would fail saying the repo is locked. Does your new system recover from such failures?

Also how do you deal with platform discrepancies? We've had in the past tests that behave radically different on different platforms, so we've also historically had the ability to record a platform specific value.

Thanks, Tamar

On Fri, Sep 1, 2017, 05:01 Jared Weakly wrote:

...

Hey y'all,

A quick ToC before I dive right in:

* What my HSOC project is on * My progress so far * Feedback welcome * What I have left to do * Theoretical potential improvements

-----------

My HSOC project was on bringing sanity to the GHC performance

test-suite.

...

My blog post on this is here: https://jaredweakly.com/blog/haskell-summer-of-code/ The Trac ticket that corresponds to this is here: https://ghc.haskell.org/trac/ghc/ticket/12758 The Phabricator ticket for this patch: https://phabricator.haskell.org/D3758

The tl;dr of my HSOC project is that GHC's performance tests currently require the programmer to add in expected numbers manually, updated them, handhold the testsuite, etc. This is a bit absurd and my project's overall aim is to reduce the effort required of the programmer to as close to zero as possible while simultaneously increasing the potential ability of the testsuite to catch regressions as much as possible.

------------

My progress so far: - I have a few comparison tools in perf_notes.py. These allow people to compare performance numbers of tests across commits - I have all the performance numbers generated by running the tests automatically stored in git notes and referenced by both the comparison tool and the testsuite - I have refactored the testsuite to use my new code that pulls expected numbers automatically from git notes (trivially passing if the note does not yet exist for that test), then it compares that expected number with the number that was gotten from running the testsuite on the latest commit. The comparison passes if it's within a certain deviation (20% by default, but can be customized by the programmer). - I have refactored all of the all.T files to use the new comparison functions for the performance tests and ensured that this doesn't break any existing tests.

------------

Anyone who wants to checkout the wip/perf-testsuite and try this out is more than welcome. Feedback on anything is welcome; comments are appreciated; discussion is welcome, etc.

-------------

What I have left to do is:

1. Finish writing up the documentation 2. Update the wiki in all the relevant places concerning additions/modifications to the testsuite and test driver 3. Make sure everyone is happy with the change (and make small changes as necessary)

--------------

Possible features and improvements I am thinking about adding in: * As a stopgap to full integration with performance tracking tools (such as Gipedia), optionally emitting a test warning with the test summary if there is any regression detected whatsoever (even if the number falls within the allowed deviation) * Some tests, such as T7702, have a somewhat nonsensical regression percentage. Ideally the testsuite could handle those better. I could potentially build in multiple ways to determine a regression (percentage, 'above a certain value', 'taking longer than X amount of time', as potential examples) * Currently some tests require installing some Haskell packages; they are skipped if the packages are not installed. I could try to build in a way to automatically attempt to install all necessary Haskell packages if someone attempts to run a test that requires them. (Perhaps using a command such as 'make test exhaustive') * The performance metric 'peak_megabytes' is sometimes not accurate enough; I could see if adding something like `RTS -h -i0.01` automatically to tests that use 'peak_megabytes' would resolve that. Currently it is a manual debugging step.

Any thoughts? Comments? Questions?

Regards, Jared Weakly _______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs