Is there any difference between the knl_7210 nodes on jlse? When I grabbed knl00-02, 08. I didn’t get the ideal performance using omni-path, the peak performance of my code was only 8.5 TFLOP/s. But when I grabbed knl03-07 using omni-path, the peak performance of my code was around 11.8 TFLOP/s. Just wondering is it caused by node difference. So when I grabbed different nodes, I can get different performance. And how to make sure to grab certain nodes on jlse? Thanks Intel MPI compiler. export I_MPI_FABRIC=tmi export I_MPI_FALLBACK=1 export KMP_AFFINITY=compact,granularity=fine mpirun -np 4 ./a.out
On Feb 7, 2017, at 12:09 AM, Hammond, Jeff R <[email protected]> wrote:
Try OSU MPI benchmarks. Those are what I use. Not perfect but easy to understand.
Jeff
Sent from my iPhone
On Feb 6, 2017, at 9:45 PM, Li, Ying <[email protected]> wrote:
yes, it is intel MPI. after export I_MPI_FABRIC=tmi and I_MPI_FALLBACK=1. Is it supposed to faster? better performance? However, my code doesn’t show much improvement. Is there any sample benchmark that I can look into? IMB-MPI1 is really confusing.
On Feb 6, 2017, at 10:45 PM, Hammond, Jeff R <[email protected]> wrote:
Intel MPI? Look up I_MPI_FABRIC docs.
Sent from my iPhone
On Feb 6, 2017, at 8:44 PM, Li, Ying <[email protected]> wrote:
Hello,
How to enable the tmi fabric on knl nodes on jlse? I tried adding “-iface ib0 -PSM” to my mpirun. But seems doesn’t work.
Best, Amy Ying Li _______________________________________________ Intel-nda mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/intel-nda