Re: [Starpu-devel] [PATCH 0/5] Merge heft and dmda.

[Cyril Roelandt <cyril.roelandt@inria.fr>]

On 12/06/2012 10:56 AM, Nathalie Furmento wrote:
> On 03/12/2012 20:46, Samuel Thibault wrote:
> > Cyril Roelandt, le Wed 28 Nov 2012 00:35:30 +0100, a écrit :
> > > The following five patches merge heft and dmda. They work well on my
> > > machine, I
> > > just haven't run any benchmark. Samuel, the last time I hacked the
> > > schedulers,
> > > you ran a benchmark on Plafrim, could you run it again and tell me if
> > > these
> > > patches affect the performance ?
> > (and the patches look good to me)
> >
> > Samuel
> >
> > _______________________________________________
> > Starpu-devel mailing list
> > Starpu-devel@lists.gforge.inria.fr
> > http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-devel
>
> The serie of patches seems to have broken StarPU.
>
> * Buildbot running with the default configuration on hannibal is
> successfull, with revision #8179
>
> http://dalton.bordeaux.inria.fr:8011/builders/hannibal-trunk-default/builds/486

On hannibal, it hangs with revision 8179.

I tried running mult/sgemm on attila, and got this output with both 
r.8233 and r.8192 (right before applying my patches):

[...]
BLAS : Program is Terminated. Because you tried to allocate too many 
memory regions.
BLAS : Program is Terminated. Because you tried to allocate too many 
memory regions.
BLAS : Program is Terminated. Because you tried to allocate too many 
memory regions.
BLAS : Program is Terminated. Because you tried to allocate too many 
memory regions.
Segmentation fault (core dumped)


> * After the commits up to revision #8197,mult/sgemm timeouts
>
> http://dalton.bordeaux.inria.fr:8011/builders/hannibal-trunk-default/builds/487
>
>
> * With the following commits up to revision #8206, many examples fail:
>
> http://dalton.bordeaux.inria.fr:8011/builders/hannibal-trunk-default/builds/488

All these failed tests use the eager scheduler, while most of my changes 
affected heft and dm*. Are you sure this issue comes from this patch 
series ? If the bug is reproducible, could you try and bisect this ?

> Did you try to run a make check on hannibal before committing?

I think so, but well, I might have forgotten.


Cyril Roelandt.