2017-01-30 10:03 GMT+01:00 Alvarez, Damian <[email protected]>:
> The pingpong latencies are a clear indicator that there is something wrong > with the MPI runtime. > Yes, obviously. > It looks to me like you are using TCP instead of InfiniBand. Did you > verify that? > No, and I can't anymore. As soon as we discussed the horrible results, I was given different tasks. Unfortunately some people here blame EasyBuild for what is probably some configuration mistake from my side, because I couldn't find the documentation how to enable OpenMPI from the foss toolchain to communicate via Infiniband correctly. I was instructed not to waste more time with these attempts, but I personally would still like to learn how to use EasyBuild efficiently. If somebody could reproduce the problem on any other Infiniband cluster, could identify the reason why the OpenMPI from the foss toolchain pingpongs more than 1000 times slower than system OpenMPI, and could come up with a clean solution to integrate a fix into EasyBuild, I hope that I can convince our team to let me give EasyBuild another try for my internship. Thank you Gunnar
