[MPICH] MPI_REDUCE with MPI_IN_PLACE fails with memory error

[Ashley Pittman]

On Wed, 2007-03-14 at 15:54 +0100, Martin Kleinschmidt wrote:
> On Mi, 14 M?r 2007, Ashley Pittman wrote:
> 
> >In our implementation we split it into pipelined 64k chunks [...]
> 
> so you are talking about a different MPI implementation? Please forgive
> my ignorance, but which implementation ist that?

It's effectively a different implementation.  We are a network vendor
and have our own MPID plugin to mpich so although the visible part of it
is MPICH the algorithms used underneath are different.  Look at
http://www.quadrics.com for details of who we are.

My point was mainly that you shouldn't lose performance by splitting it
into smaller chunks, even without pipelineing (which is considerably
harder to do) you may well actually get better performance.

Ashley,