Re: [Starpu-devel] starpu_data_unregister_submit() and MPI

[Benoît Lizé <benoit.lize@gmail.com>]

An additional observation: on my machine, removing the starpu_data_unregister_submit() call solves the issue.

On Wed, Mar 12, 2014 at 5:53 PM, Benoît Lizé <benoit.lize@gmail.com> wrote:

Hello,

Here is a small example that doesn't work. It is stuck when the node count is not 1. I've tested it with OpenMPI 1.4.4 (very old, sorry about that) on debian testing with StarPU r2389. It can be run with:

mpirun -np 2 xterm -geometry 160x100 -e gdb -ex run ./test 

It is however not stuck at the same step as my full program, I don't know why. Of course it is entirely possible that I missed something elsewhere.

On Wed, Mar 12, 2014 at 5:14 PM, Benoît Lizé <benoit.lize@gmail.com> wrote:

I haven't found the cause either.

I will try to isolate a small test case that exhibits the issue and get back to you.

On Wed, Mar 12, 2014 at 4:58 PM, Samuel Thibault <samuel.thibault@ens-lyon.org> wrote:

Benoît Lizé, le Wed 12 Mar 2014 11:02:03 +0100, a écrit :

> Here it is (with STARPU_NCPUS=1):

Ok, I don't have any idea by just looking at the backtrace.

Nathalie, was there any reason for calling the unregister hook so late?
It should rather go as soon as the beginning actually, so that the hook
operations are properly waited for by the rest of
_starpu_data_unregister.  I don't think it is the bug at stake, however.

Benoît, would you have a simple example that triggers the bug?

Samuel