Re: [Starpu-devel] StarPU questions

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

David Pereira, le Thu 25 Sep 2014 23:37:07 +0000, a écrit :
> I have executed my StarPU version on two GPUs (without CPUs) and on two 
> GPUs +
> a dual socket CPU (2 NUMA nodes of 6 cores each), and the former is faster 
> for
> every scheduler that I have used (peager, pheft, ws, dm and dmda). Is this
> supposed to be normal? I mean, the difference is quite expressive even 
> though I
> don't have a high level of paralellism, I thought that StarPU would try to
> minimize the global execution time (and I suppose that it would be by 
> executing
> most of the tasks on both GPUs). However, StarPU schedules a significant 
> number
> of tasks on the CPU and this degrades overall performance. I think that this
> may be normal for the "peager" and "ws" schedulers but not for the other 
> ones.

Well, scheduling algorithms are never perfect. The dmda algorithms, for
instance, don't actually minimize global execution time (which is NP complete
anyway), but the local execution time of each task. Setting priorities on 
tasks
can make a big difference, to make sure that critical tasks get scheduled 
first,
and thus sensibly, and scheduled the rest after that.

> Also, I'm thinking of trying to use the Xeon Phi. The trunk documentation
> explains just a little bit how I can work with it. My question is: Does 
> StarPU
> sees the MIC as a single worker or each of its cores as a single worker?.Or 
> is
> it just like a normal CPU where I can create combined workers?

ATM it is just like a normal socket where you can create combined workers.
We have thought about exposing it as a single worker, or perhaps just a few
workers, but nobody worked on implementing that so far.

Samuel