[Mpi3-rma] Updated Proposal 1

Underwood, Keith D keith.d.underwood at intel.com
Wed Dec 8 10:38:52 CST 2010


> > (b) Do we know any network (except ones that emulate one-sided
> > communication over two-sided messaging) that provide strict ordering?
> As
> > far as I know, most networks only provide ordering between PUTs (or
> > atomic PUTs). Ordering between GETs is also not provided, unless the
> > user does a flush after each GET operation. Is this only for
> convenience?
> I don't know. Just that you're aware: the current default would force
> the MPI implementation to flush after each get. However, I am also
> planning to add an advice to users somewhere *very visible* that the
> user should really strive to provide these info arguments if possible
> (because the performance difference could be hideous). So the editing
> is
> not finished but 11.8.2 is ready for discussion ;-).

Yeah, I think we underspecified "partial".  Accumulate->GetAccumulate should be ordered, but GetAccumulate->Accumulate should not.  The former is trivial if we have Accumulate->Accumulate ordering already.  As an example, InfiniBand gives the former, but not the latter (unless you set the fence bit and then ugly things happen).

Keith




More information about the mpiwg-rma mailing list