Re: [Starpu-devel] Incoherence between starpu_mpi_init and starpu_mpi_init_conf

[Samuel Thibault <samuel.thibault@inria.fr>]

Philippe SWARTVAGHER, le mer. 12 févr. 2020 14:06:24 +0100, a ecrit:
> Le 12/02/2020 à 14:00, Samuel Thibault a écrit :
> > No, it is the original semantic of starpu_mpi_init(). Initially you had
> > to call starpu_init() then starpu_mpi_init(). We added
> > starpu_mpi_init_conf() as a helper to do both. Perhaps the doc could be
> > reworded to make it clearer.
> 
> My point is: if you call starpu_init() and then starpu_mpi_init(), then
> _mpi_backend._starpu_mpi_backend_init(conf) and eventual core reservation
> for comm progression, are not executed ! Indeed, these instructions are only
> in starpu_mpi_init_conf().

Ah, by backend you meant nmad vs mpi backend, I understood libstarpu.

Concerning CPU reservation, starpu_init+starpu_mpi_init can't do it
since it's a parameter passed to starpu_init.

Concerning calling _starpu_mpi_backend_init, starpu_mpi_init should be
doing it, yes.

Samuel