Saliya,
from ompi/mca/coll/sm/coll_sm_module.c in mca_coll_sm_comm_query()
sm_module->super.coll_allgatherv = NULL;
that means the coll sm module does *not* implement allgatherv, so
openmpi will use the next module
(which is very likely the default module, that is why there is no
performance improvement in your specific benchmark)
Cheers,
Gilles
On 12/10/2015 2:53 AM, Saliya Ekanayake wrote:
Hi,
In a previous email, I wanted to know how to enable shared memory
collectives and I was told setting the coll_sm_priority to anything
over 30 should do it.
I tested this for a microbenchmark on allgatherv, but it didn't
improve performance over the default setting. See below, where I
tested for different number of processes per node on 48 nodes. The
total message size is kept constant at 2400000 bytes (or 2.28MB).
Am I doing something wrong here?
Thank you,
saliya
Inline image 1
--
Saliya Ekanayake
Ph.D. Candidate | Research Assistant
School of Informatics and Computing | Digital Science Center
Indiana University, Bloomington
Cell 812-391-4914
http://saliya.org
_______________________________________________
users mailing list
us...@open-mpi.org
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
Link to this post:
http://www.open-mpi.org/community/lists/users/2015/12/28153.php
