Attempt at a real world benchmark

Bardur Arantsson spam at scientician.net
Fri Dec 9 07:31:31 UTC 2016


On 2016-12-08 17:04, Joachim Breitner wrote:
> Hi,
> 
> Am Donnerstag, den 08.12.2016, 01:03 -0500 schrieb Joachim Breitner:
>> I am not sure how useful this is going to be:
>>  + Tests lots of common and important real-world libraries.
>>  − Takes a lot of time to compile, includes CPP macros and C code.
>> (More details in the README linked above).
> 
> another problem with the approach of taking modern real-world code:
> It uses a lot of non-boot libraries that are quite compiler-close and
> do low-level stuff (e.g. using Template Haskell, or stuff like the). If
> we add that not nofib, we’d have to maintain its compatibility with GHC
> as we continue developing GHC, probably using lots of CPP. This was
> less an issue with the Haskell98 code in nofib.
> 
> But is there a way to test realistic modern code without running into
> this problem?
> 

This may be a totally crazy idea, but has any thought been given a
"Phone Home"-type model?

Very simplistic approach:

  a) Before it compiles, GHC computes a hash of the file.
  b) GHC has internal profiling "markers" in its compilation pipeline.
  c) GHC sends those "markers" + hash to some semi-centralized
highly-available service somewhere under *.haskell.org.

The idea is that the fact that "hashes are equal" => "performance should
be comparable". Ideally, it'd probably be best to be able to have the
full source, but that may be a tougher sell, obviously.

(Obviously would have to be opt-in, either way.)

There are a few obvious problems with this, but an obvious win would be
that it could be done on a massively *decentralized* scale. Most
problematic part might be that it wouldn't be able to track things like
"I changed $this_line and now it compiles twice as slow".

Actually, now that I think about it: What about if this were integrated
into the Cabal infrastructure? If I specify "upload-perf-numbers: True"
in my .cabal file, any project on (e.g.) GitHub that wanted to opt-in
could do so, they could build using Travis, and voila!

What do you think? Totally crazy, or could it be workable?

Regards,



More information about the ghc-devs mailing list