[mpich-discuss] MPI_Allreduce is slow for 7 processes or more

Ken Raffenetti raffenet at mcs.anl.gov
Tue May 5 02:08:46 CDT 2015

On 05/05/2015 05:03 AM, David Froger wrote:
> I did not tried mpich3.2b2 yet (will do tomorrow), but I've ask another
> question here, just on the time to execute MPI_Barrier:
> http://stackoverflow.com/questions/30039802/what-control-mpi-barrier-time-to-execute
> Seems that that was why `bench_mpi.cxx` was slow.
The underlying issue is that recent MPICH causes processes to busy poll 
for incoming messages. In over-subscription scenarios, this can have 
significant performance implications. There is a ongoing work to address 
the busy polling issue for MPICH 3.3. In the interim, our advice is to 
not oversubscribe processes for the available hardware resources.

discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:

More information about the discuss mailing list