Hi Thomas, as far as I know MPI does _not_ guarantee asynchronous progress (unlike OpenSHMEM) because it would require some implementations to start a progress thread.
Jeff has a nice blog post regarding this: http://blogs.cisco.com/performance/mpi-progress I was surprised to see this behavior in my microbenchmark as I thought Infiband would provide good hardware support for rDMA (for some time) and this issue was "solved" in the sense "if you have a modern network, don't worry about it". I don't think it's unique to Open MPI, as I noticed that MPICH-based implementations seem to behave similarly (without have looked into their actual sources). It seems the MVAPICH people are working heavily on this, as I see their papers pop up once in a while on this topic, though. Sadly, this was a showstopper for colleagues of mine checking out MPI RMA for their algorithm implementation, as this progress-related waiting time was just too high. We are looking into a p2p-based solution now (although RMA would have been more elegant). Cheers, Marc-Andre On 30.06.15 11:14, Thomas Jahns wrote: > On 06/29/15 17:25, Nathan Hjelm wrote: >> This is not a configuration issue. On 1.8.x and master we use two-sided >> communication to emulation one-sided. Since we do not currently have >> async progress this requires the target to call into MPI to progress >> RMA >> communication. >> >> This will change in 2.x. I will be adding a new component that does >> real RMA on supported networks (ib, gemini, aries). At some point we >> will also have support for async progress which will enable >> passive-target on other networks. > > But isn't aynch progress required for MPI 2.x conformance? The Open > MPI 1.6.5 headers I have #define MPI_VERSION to 2. The 1.8.4 headers I > have even define it to 3. Why isn't a background thread used on > platforms that don't have supported networks? > > Thomas > > > > _______________________________________________ > 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/2015/06/27227.php > -- Marc-Andre Hermanns Jülich Aachen Research Alliance, High Performance Computing (JARA-HPC) German Research School for Simulation Sciences GmbH Schinkelstrasse 2 52062 Aachen Germany Phone: +49 2461 61 2509 Fax: +49 241 80 6 99753 www.grs-sim.de/parallel email: m.a.herma...@grs-sim.de
smime.p7s
Description: S/MIME Cryptographic Signature
