[Mpi3-rma] RMA proposal 1 update

Jeff Hammond jeff.science at gmail.com
Fri May 21 17:54:24 CDT 2010

> So, why not just use active target w/ MPI_Win_fence()?

Many parts of NWChem could probably be rewritten to do exactly that.
However, NWChem is not the only code which uses GA, although clearly
the most important one that we know of.

It is exactly because of the efforts to reconfigure NWChem to not use
GA_Sync everywhere and because of efforts to develop better quantum
chemistry codes that I want all three flush commands.

I know I sound schizoid sometimes.  I want allflushall for backwards
compatibility and for the relatively rare cases when I really need to
sync across an entire window collectively.

It is not clear to me that I can mix Win_fence with Lock/Unlock, hence
it is very hard to implement GA without some other form of collective


Jeff Hammond
Argonne Leadership Computing Facility
jhammond at mcs.anl.gov / (630) 252-5381

More information about the mpiwg-rma mailing list