Re: [Starpu-devel] [Starpu-commits] r8070 - in trunk/mpi: examples examples/cholesky examples/complex examples/mpi_lu examples/stencil include src tests

[Nathalie Furmento <nathalie.furmento@labri.fr>]

On Nov 22, 21:40, Cyril Roelandt wrote:
> On 11/22/2012 08:26 PM, nfurmento@users.gforge.inria.fr wrote:
> >Author: nfurmento
> >Date: 2012-11-22 20:26:20 +0100 (Thu, 22 Nov 2012)
> >New Revision: 8070
> >
> >Log:
> >mpi: add a parameter initialize_mpi to starpu_mpi_init, then we no longer 
> >rely on MPI_Initialized() to detect if MPI is already initialised or not, 
> >we just ask the upper layer to tell us if it needs to be initialised
> >
> 
> 
> Why is it bad to rely on MPI_Initialized?

I did not use the word bad ...

Samuel tried today to use SMPI, the MPI layer of Simgrid. It did not
work as it does not provide a implementation of MPI_Initialized(). It
could have been added, but Samuel said it was not so easy.

It is also not thread safe. There is always the possibility one thread
calls MPI_Init and another one starpu_mpi_init ...

Nathalie