Fabian Wein <fabian.w...@fau.de> writes: >> Am 30.10.2015 um 21:45 schrieb Jeff Squyres (jsquyres) <jsquy...@cisco.com>: >> >> Oh, that's an interesting idea: perhaps the "bind to numa" is >> failing -- but perhaps "bind to socket" would work. >> >> Can you try: >> >> /opt/openmpi-1.10.0-gcc/bin/mpiexec -bind-to numa -n 4 hostname >> /opt/openmpi-1.10.0-gcc/bin/mpiexec -bind-to socket -n 4 hostname >> > Both report the same error. Interestingly -bind-to-socket works
There's something badly wrong with your build or installation if -bind-to socket isn't equivalent to --bind-to-socket (which is the default, as you should see if you run hwloc-ps under mpirun). I just built 1.10 on Ubuntu 14.04 against the native libhwloc 1.8 (without libnuma1-dev). It works as expected on that single-socket system and --bind-to numa fails as there's no numanode level. > but it does not bring me the performance I expect for the petsc benchmark. Without a sane installation it's probably irrelevant, but performance relative to what? Anyhow, why don't you want to bind to cores, or at least L2 cache, if that's shared?
