[Mpi-forum] Persistence Working Group Proposed

Richard Treumann treumann at us.ibm.com
Wed Feb 25 12:11:30 CST 2009


In my opinion the lack of enthusiasm for point to point persistent
operations does not tell us much about the usefulness of persistent
operations in general. When the MPI-1 Forum decided an MPI_Isend_init or
MPI_Irecv_init must be a local operation and that an MPI_Start of a
persistent send on one side must match any form of receive (persistent or
not) on the other (and vice versa) most of the opportunity for optimization
was lost.

For a persistent collective to be fully optimized, the MPI library probably
must be able to do collective communication at the CC_INIT and return to
each task a persistent request handle that is irrevocably bound to all the
handles created at the other tasks by this collective operation.   It
should not be valid to create two persistent BCAST requests at the root
with different send buffers and one persistent BCAST at the non-roots just
so root could BCAST from either buffer.  (even if somebody thinks that
would be a cool bit of generality)

Dick Treumann  -  MPI Team
IBM Systems & Technology Group
Dept X2ZA / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846         Fax (845) 433-8363
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpi-forum/attachments/20090225/3b60c461/attachment-0001.html>

More information about the mpi-forum mailing list