On Fri, Jun 26, 2009 at 6:24 PM, Dominic Walsh<dwa...@abingdon.oilfield.slb.com> wrote: > I have just been trying out SCALI MPI with Valgrind and have observed that > when running across multiple nodes the process seem to end up spinning in > MPI_Init. Smaller problems running with shared memory appear fine. > > The configuration is: > > 8 Nodes connected by Infiniband > 2 Sockets/Node with 1 MPI process each > Plenty of spare memory > No competing processes – memcheck is at 100% > > My hunch is there is weird packet dropping effect but hoping someone out > there has an idea. Was hoping I could tell Valgrind to skip instrumenting > libmpi.so? Either I didn’t read the manual well enough or it breaks a > fundamental principle (which I suspect it might)
Have you noticed the thread about Valgrind and memcpy() on the ofa-general mailing list ? See also http://lists.openfabrics.org/pipermail/general/2009-June/060037.html. Bart. ------------------------------------------------------------------------------ _______________________________________________ Valgrind-users mailing list Valgrind-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/valgrind-users