Optimizing the Performance of SPEC MPI with KNL and Omni-Path
Optimizing the Performance of SPEC MPI with KNL and Omni-Path
Tuning the Eager threshold has a significant impact on application performance by avoiding the synchronization of rendezvous protocol and thus yielding better communication computation overlap.
Library Version: MVAPICH2 2.3rc1
Runtime Flags:
| milc | MV2_SMP_EAGERSIZE=256000 |
| leslie3d | MV2_SMP_EAGERSIZE=1024000 |
| pop2 | None |
| lammps | MV2_SMP_EAGERSIZE=512000 |
| wrf2 | MV2_SMP_EAGERSIZE=512000 |
| tera_tf | MV2_SMP_EAGERSIZE=128000 |
| lu | MV2_SMP_EAGERSIZE=1024000 |
System Details:
| CPU Model | Intel Xeon Phi 7250 |
| CPU Core Info | 68 @ 1.4 GHz |
| IB Card | 100Gb/sec Intel Omni-Path (OPA) |
| IB Switch | Full Fat Tree 100Gb/sec Intel Omni-Path with 6 core switches |
Submitted by Amit Ruhela @ OSU
Last Modified March 6, 2018, 1:54 p.m.
