PSTSWM Communication Characterization

Performance Studies using

PSTSWM

Best Point-to-Point Message-Passing Performance for Each Platform

These results represent the highest measured bandwidths for each platform for the single iteration, nonoverlap communication protocols for internode communication. That is, in a system comprised of a cluster of SMP nodes, bandwidth is measured when one processor in one SMP node is communicating with one processor in a different SMP node. The one exception to this rule is that the SPP meaurements are for communication within a hypernode, in which all processors have access to a global shared memory.

Results are presented for bidirectional bandwidth using MPI, bidirectional bandwidth using any communication library, unidirectional bandwidth using MPI, and unidirectional bandwidth using any communication library, The legends list the platforms in order of their performance.

Bidirectional Bandwidth (MPI)

Bidirectional Bandwidth (Optimal)

Unidirectional Bandwidth (MPI)

Unidirectional Bandwidth (Optimal)

Patrick H. Worley / ( worleyph@ornl.gov)
Last Modified Monday, 15-Jul-2002 09:58:36 EDT.
5097 accesses since 1/2/96.