Accéder au contenu.
Menu Sympa

starpu-devel - Re: [Starpu-devel] StarPU overhead for MPI

Objet : Developers list for StarPU

Archives de la liste

Re: [Starpu-devel] StarPU overhead for MPI


Chronologique Discussions 
  • From: Samuel Thibault <samuel.thibault@inria.fr>
  • To: Philippe SWARTVAGHER <philippe.swartvagher@inria.fr>
  • Cc: starpu-devel@lists.gforge.inria.fr, Alexandre Denis <alexandre.denis@inria.fr>
  • Subject: Re: [Starpu-devel] StarPU overhead for MPI
  • Date: Wed, 5 Feb 2020 18:03:23 +0100
  • 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>
  • Organization: I am not organized

Hello,

Philippe SWARTVAGHER, le mer. 29 janv. 2020 17:19:10 +0100, a ecrit:
> The overhead seems a little bit huge (a small comm lasts 10 times longer).
> Do you know what could be the origin of this overhead ?

StarPU tasks (here, one is used for each MPI request) are costly, yes,
because they track a data dependency, check that it's not on a GPU, etc.
so a few additional µs is not really surprising. It is a bit surprising
that it's so many µs, a gprof would help to get an idea where most time
is spent.

Samuel




Archives gérées par MHonArc 2.6.19+.

Haut de le page