Re: [Starpu-devel] Issue with distributed NUMA-aware StarPU and dmda scheduler

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

Samuel Thibault, le ven. 13 mars 2020 17:55:39 +0100, a ecrit:
> Philippe SWARTVAGHER, le ven. 13 mars 2020 14:02:38 +0100, a ecrit:
> > #5  0x00007ffff7fb0053 in _starpu_mpi_handle_request_termination 
> > (req=req@entry
> > =0x55555581b640) at ../../../mpi/src/mpi/starpu_mpi_mpi.c:861
> 
> It seems that it frees the packed data with free(), while the documentation 
> says
> that the pack_data method is supposed to use starpu_malloc_on_node_flags
> to allocate it. Please try the attached patch to properly release it.

Yes, that was bogus (and the patch was bogus too), I pushed and
backported a fix.

Samuel

> diff --git a/mpi/src/mpi/starpu_mpi_mpi.c b/mpi/src/mpi/starpu_mpi_mpi.c
> index 01c46fe30..1548641d6 100644
> --- a/mpi/src/mpi/starpu_mpi_mpi.c
> +++ b/mpi/src/mpi/starpu_mpi_mpi.c
> @@ -858,7 +858,7 @@ static void 
> _starpu_mpi_handle_request_termination(struct _starpu_mpi_req *req)
>                                       int ret;
>                                       ret = 
> MPI_Wait(&req->backend->size_req, MPI_STATUS_IGNORE);
>                                       STARPU_MPI_ASSERT_MSG(ret == 
> MPI_SUCCESS, "MPI_Wait returning %s", _starpu_mpi_get_mpi_error_code(ret));
> -                                     free(req->ptr);
> +                                     starpu_free_on_node(STARPU_MAIN_RAM, 
> req->ptr, req->backend->envelope->size);
>                                       req->ptr = NULL;
>                               }
>                               else if (req->request_type == RECV_REQ)