Re: [Starpu-devel] Performance issues when specifying const and __attribute__ ((output))

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Mehdi AMINI, le Mon 02 Jul 2012 14:26:05 +0200, a écrit :
> Le 02/07/2012 13:52, Samuel Thibault a écrit :
> >Mehdi AMINI, le Mon 02 Jul 2012 12:10:19 +0200, a écrit :
> >>$ STARPU_NCUDA=1 STARPU_NCPUS=0 ./bicg_starpu_naive
> >>192.2
> >>$ STARPU_NCUDA=1 STARPU_NCPUS=0 ./bicg_starpu
> >>746.5
> >
> >Ouch!
> >
> >Could you dump traces so we have an idea of what is happening?
> >
> 
> Are the FXT traces enough? You'll find them here for both versions :
> 
> http://mehdi.amini.fr/StarPU/

So it seems it spends almost all its time in PushingOutput, which is not
supposed to happen. Could you perhaps run in gdb, interrupt it during
the run, and use

thread apply all bt

to have an idea at where it most often is? Or alternatively, compile
with gprof, but that might be more involved.

Would it be possible to see your code?  It is hard to have an idea
without being able to run it.

Samuel