Accéder au contenu.
Menu Sympa

starpu-devel - [Starpu-devel] Why starpu_mpi_barrier waits for all taks ?

Objet : Developers list for StarPU

Archives de la liste

[Starpu-devel] Why starpu_mpi_barrier waits for all taks ?


Chronologique Discussions 
  • From: Philippe SWARTVAGHER <philippe.swartvagher@inria.fr>
  • To: starpu-devel@lists.gforge.inria.fr
  • Subject: [Starpu-devel] Why starpu_mpi_barrier waits for all taks ?
  • Date: Tue, 14 Apr 2020 18:02:00 +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>

Hello,

I'm trying to fix a bug reported by Nathalie: mpi/tests/sendrecv_gemm_bench isn't working with MPI backend because of calls to starpu_mpi_barrier().

Indeed, the MPI version of starpu_mpi_barrier() calls starpu_task_wait_for_all(). Hence the deadlock in the test:

- starpu_pause()

- task submission

- starpu_mpi_barrier() <-- cannot finish since starpu_task_wait_for_all is blocking

- starpu_release()


Why starpu_mpi_barrier waits for all tasks ? Is it because with MPI we cannot have concurrent barriers ?

Could we consider creating a function starpu_mpi_barrier_unsafe(), calling directly MPI_Barrier ?

--
Philippe SWARTVAGHER

Doctorant
Équipe TADaaM, Inria Bordeaux Sud-Ouest





Archives gérées par MHonArc 2.6.19+.

Haut de le page