[PATCH] Add benchmarking-only mode to the test suite

Hal Finkel hfinkel at anl.gov
Thu Jun 5 02:37:14 PDT 2014


----- Original Message -----
> From: "Yi Kong" <kongy.dev at gmail.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "Tobias Grosser" <tobias at grosser.es>, llvm-commits at cs.uiuc.edu
> Sent: Wednesday, June 4, 2014 4:13:47 AM
> Subject: Re: [PATCH] Add benchmarking-only mode to the test suite
> 
> Hi Hal Finkel,
> 
> I've been running the benchmarking mode for almost two weeks and
> didn't seem to miss any performance regression compared to full run.
> I
> think you should commit the patch.

r210251, thanks! I did not include the polybench exclusions, have you been using those as well? I'm not sure what the final verdict on those was.

 -Hal

> 
> -Yi
> 
> On 23 May 2014 18:22, Yi Kong <kongy.dev at gmail.com> wrote:
> > On 23 May 2014 17:38, Tobias Grosser <tobias at grosser.es> wrote:
> >> So what do these numbers tell us? We have a lot more cache misses
> >> in the
> >> first run. however, the ratio (66%) remains identical. Does this
> >> mean we
> >> load values in generally a lot more often? That seems surprising.
> >
> > L1 cache misses count varies greatly, which means once L1d cache is
> > full, the behaviour becomes really unpredictable.
> >
> >>
> >>> Hal, can you add -DSMALL_DATASET to Makefile or change the array
> >>> size
> >>> for STANDARD_DATASET to avoid filling up L1d cache?
> >>
> >>
> >> The interesting optimization problem here is exactly cache
> >> behavior. By
> >> using -DSMALL_DATASET we just avoid it and can measure it any
> >> more. That
> >> would be unfortunate.
> >
> > Otherwise they produce lots of false positives, especially on
> > systems
> > with small cache. We could revert it once Polly is capable of
> > optimizing these code to a decent state for benchmarking.
> >
> >> Cheers,
> >> Tobias
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory



More information about the llvm-commits mailing list