Accéder au contenu.
Menu Sympa

starpu-devel - Re: [Starpu-devel] OpenMP

Objet : Developers list for StarPU

Archives de la liste

Re: [Starpu-devel] OpenMP


Chronologique Discussions 
  • From: Samuel Thibault <samuel.thibault@ens-lyon.org>
  • To: Nathalie Furmento <nathalie.furmento@labri.fr>
  • Cc: uma boregowda <umaboregowda@gmail.com>, "starpu-devel@lists.gforge.inria.fr" <starpu-devel@lists.gforge.inria.fr>
  • Subject: Re: [Starpu-devel] OpenMP
  • Date: Fri, 30 Mar 2012 13:07:27 +0200
  • List-archive: <http://lists.gforge.inria.fr/pipermail/starpu-devel>
  • List-id: "Developers list. For discussion of new features, code changes, etc." <starpu-devel.lists.gforge.inria.fr>

Nathalie Furmento, le Fri 30 Mar 2012 13:00:38 +0200, a écrit :
> On 30/03/2012 12:02, uma boregowda wrote:
> Well, OpenMP programs alone can execute parallel tasks already. What
> StarPU brings is choosing for each task whether it should run on a
> GPU
> (that requires writing the CUDA or OpenCL kernel), or on a bunch of
> CPUs, using the OpenMP parallel implementation.
>
>
> It means, that means starpu will not do anything more to thread
> scheduling, if
>
> one is using only CPU to run the applciation with threads.
>
>
> As Samuel said before, threads are not scheduled by StarPU.

Note: for the threads inside parallel tasks. Tasks are still scheduled
by StarPU.

On a machine with a lot of cores, and for an application whose structure
is not just flat, StarPU can be useful because scalability of the OpenMP
implementation may be limited and so in that case you can not just run
the OpenMP implementation on all cores, but rather schedule parallel tasks
concurrently. Nested parallelism in OpenMP also permits that, but it is
only tractable for very regular applications. For unstructured
applications, StarPU can provide the scheduling of parallel tasks over
sets of cores, to both benefit from the OpenMP parallel implementation
inside parallel tasks, and proper scheduling of unstructure application
over the big machine.

Samuel





Archives gérées par MHonArc 2.6.19+.

Haut de le page