6 Apr
2016
6 Apr
'16
5:47 p.m.
I was seeing odd results from the kmi_hash benchmark which heavily uses the MPI_Test interface. KNL was much slower than a BGQ node. Looking at a standalone test case of MPI_Test, KNL with Intel MPI is much slower to execute MPI_Test than bgq. results for 64 ranks: bgq: loops: 5000000 time: 9.306008 test/s: 537287.314173 knl: loops: 5000000 time: 78.315594 test/s: 63844.245215 I'm guessing most codes don't use MPI_Test but we should probably figure out what's going on. kevin