[Mpi3-hybridpm] endpoint proposal [helper threads proposal updated]

Douglas Miller dougmill at us.ibm.com
Thu May 20 07:01:44 CDT 2010

I have also attached an updated version of the Helper Threads proposal to
Ticket #217 , to which there is a link from the MPI3 Hybrid Programming
wiki. This includes some performance graphs for the BG/P Collective network
showing its need for multiple threads.

Douglas Miller                  BlueGene Messaging Development
IBM Corp., Rochester, MN USA                     Bldg 030-2 A410
dougmill at us.ibm.com               Douglas Miller/Rochester/IBM

             "Snir, Marc"                                                  
             <snir at illinois.ed                                             
             u>                                                         To 
             Sent by:                  "mpi3-hybridpm at lists.mpi-forum.org" 
             mpi3-hybridpm-bou         <mpi3-hybridpm at lists.mpi-forum.org> 
             nces at lists.mpi-fo                                          cc 
                                       [Mpi3-hybridpm] endpoint proposal   
             05/19/2010 06:55                                              
             Please respond to                                             
             mpi3-hybridpm at lis                                             

I attached a new version of the proposal on the wiki: main.2.pdf<

Main changes:

Changed the initialization functions as discussed in Chicago

Added the COMM_MERGE functionality we discussed (page 20-23)

Got rid of the new header files for OpenMP and PGAS

Points to discuss:

Page 10
The issue of threads being attached by default to endpoint 0 for backward
compatibility. Need to discuss the current solution (an extra
initialization argument) vs. the option of having all threads attached by
default to endpoint 0.

The issue of common level of thread support: common to all endpoints, to
all endpoints at a process, different for each endpoint.

Page 12 -- do we want cached attributes with endpoints?

Page 20 -- do we want MPI_COMM_MERGE? Useful but takes a while to explain

Page 35 & 37 - PGAS initialization (we never went over this text)

Mpi3-hybridpm mailing list
Mpi3-hybridpm at lists.mpi-forum.org

More information about the mpiwg-hybridpm mailing list