Performance Studies using PSTSWM

SGI/Cray Research T3E-900 Protocol Performance

(MPI/SHMEM Comparison - May 11, 1998)

The SGI/Cray Research T3E-900 is a distributed-memory parallel architecture built around a high-performance 3D torus interconnect. We looked at the protocol sensitivities of two different message-passing libraries: MPI and an application-specific library built using the SHMEM one-sided communication operations get and put.

The MPI and SHMEM communication protocols are not directly comparable, but some general conclusions can be drawn from the protocol sensitivity experiments. The primary conclusion is that there are few similarities between the MPI and SHMEM results:

• Most of the optimal MPI protocols do not even exist in the SHMEM implementation.
• The MPI results show much more performance sensitivity than the SHMEM results.
• The MPI results show strong similarities in performance sensitivity between sets of algorithms. Corresponding similarities do not hold for the SHMEM data.
• Overlap protocols are often optimal when using SHMEM, but are less useful when using MPI.

The only significant agreement between the MPI and SHMEM results is that that Experiments B and C show less performance variation than Experiment A. There is also some similarity in how the performance variation varies - proprotional to the number of processors (1D) and inversely proprotional to the problem granularity. This is consistent with contention not being an issue.

PSTSWM Performance Page

---

Patrick H. Worley / ( worleyph@ornl.gov)
Last Modified Monday, 15-Jul-2002 10:25:17 EDT.
85980 accesses since 1/2/96.