Re: [Starpu-devel] Performance decline from StarPU 1.2.3 to 1.2.4

[Mirko Myllykoski <mirkom@cs.umu.se>]

Hi,

Thank you for taking look at the problem.

I modified StarPU 1.2.4 such that CHUNKS_NFREE was set to 1. No change. 
StarPU 1.2.3 gives a median runtime of 1635 MS and StarPU 1.2.4 + 
CHUNKS_NFREE set to 1 gives a median runtime of 2341 MS.

It also looks like StarPU 1.2.4 is struggling to calibrate the 
performance models. In the end, I get modelled_task_cnt = 695 (94.7%) 
but it seems to take more calibration runs.

I also experimented with STARPU_REGRESSION_BASED but it did not change 
anything.

 - Mirko

On 2018-09-27 15:45, Samuel Thibault wrote:
> Hello,
>
> Mirko Myllykoski, le jeu. 27 sept. 2018 15:21:01 +0200, a ecrit:
> > noted that the performance worsens
> > significantly when I move from StarPU 1.2.3 to StarPU 1.2.4. I appears 
> > that
> > the performance of StarPU 1.3.xx is similar to StarPU 1.2.4.
> >
> > Has something significant changed in StarPU 1.2.4?
>
> Not big changes, but there can be subttle changes which could make a
> difference. Ideally using a git bisect would pinpoint the culprit
> change.
>
> I notably see in the traces that the Allocating and Freeing states for
> the main ram are way longer in 1.2.4, I don't see any reason why this
> should be so, malloc.c doesn't have really changed so much. There is
> just one thing: in src/datawizard/malloc.c, CHUNKS_NFREE was set to 4,
> while it was 1 in StarPU 1.2.3, you could check whether setting it back
> to 1 fixes the issue.
>
> Thanks,
> Samuel