Andreas,

Thanks again for you time with this.

On Mon, Feb 14, 2011 at 12:40 AM, Andreas Kloeckner
<[email protected]> wrote:
>
> I can't reproduce your results using mpich2 version 1.3.2. If I mpirun
> -n 2 your program, both processes get something in their stdout. Perhaps
> a bug in your mpich2 version?

That was my initial thought too. I tested the repro code using the
same PyCUDA/mpi4py/mpich2 versions used on the Centos 5.2 cluster on a
Ubuntu 9.04, and it works. This leaves only Python version (2.4.3
versus 2.6.2) as the obvious different. On the Centod cluster, basic
forking tests in Python all work as expected, as does stdout routing
of Python scripts from mpich2. As best as I can tell each mechanism
needed works correctly separately, but not together. After playing
around with strace a bit, I have almost convinced myself  this is a
bug in the Python 2.4.3 standard library, but that is only a hunch.

I will try building the same version of mpich2 you tested and see if
anything changes, but I am not holding my breath.

>
> In the meantime, I've integrated your earlier hack into PyCUDA. If the
> case is encountered, a warning is also issued.
>

Thanks for that.

_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda

Reply via email to