Open MPI's OpenFabrics support will spawn up to two additional blocking threads (they wait for asynchronous verbs events of various flavors). They consume a few resources, but typically are not used much. They don't cause any noticeable change in performance.
On May 17, 2010, at 12:38 PM, Ping Luo wrote: > > Hello, > > I found when running am MPI program that is linked against OpenMPI library, > for each MPI task, OpenMPI will spawn three threads, as the sample shown > below: > > > $ ps axms > ... > 13536 3565 0000000000000000 - - > - - pts/14 0:00 mpirun -n 2 ./a.out > 13536 - 0000000000000000 0000000000000000 0000000000000000 > 0000000180015ee2 S+ - 0:00 - > 13536 3566 0000000000000000 - - > - - pts/14 0:00 ./a.out > 13536 - 0000000000000000 0000000000010000 0000000000000006 > 00000001800104e0 SLl+ - 0:00 - > 13536 - 0000000000000000 0000000000000000 0000000000000006 > 00000001800104e0 SLl+ - 0:00 - > 13536 - 0000000000000000 0000000000000000 0000000000000006 > 00000001800104e0 SLl+ - 0:00 - > 13536 3567 0000000000000000 - - > - - pts/14 0:00 ./a.out > 13536 - 0000000000000000 0000000000010000 0000000000000006 > 00000001800104e0 SLl+ - 0:00 - > 13536 - 0000000000000000 0000000000000000 0000000000000006 > 00000001800104e0 SLl+ - 0:00 - > 13536 - 0000000000000000 0000000000000000 0000000000000006 > 00000001800104e0 SLl+ - 0:00 - > ... > > > Does OpenMPI always spawn three threads for each MPI task by default? Can the > number > of threads be changed? Why is it implemented this way? Is there any case that > it will > degrade performance because of over-subscription? > > Thanks, > > -- > Ping Luo > 105 CSA, Building #517 > Texas A&M University > Email: ping...@sc.tamu.edu > Phone: (979)862-3107 > _______________________________________________ > users mailing list > us...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/users > -- Jeff Squyres jsquy...@cisco.com For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
