next up previous contents
Next: Algorithm Performance on Up: Empirical Results --- Previous: Empirical Results ---

Algorithm Performance on an IBM SP1 and SP2

We have ported this software to the IBM SP1 and SP2. The available communications on the IBM parallel machines required the use of the non-blocking buffered communications paradigm, because active messages are not implemented on this hardware. We chose to use the Message Passing Interface (MPI) for the communications language, because it is being developed as a communications standard for multi-processors with strong emphasis on optimizing message-passing performance. In table 7, we present empirical performance data for the IBM SP1 using MPI and the IBM SP2 using standard Transmission Control Protocol (TCP)/Internet Protocol (IP) based communications through the embedded communications switch. This preliminary data from the IBM scalable parallel processors (SPPs), based on workstation clusters with switched network communications, shows that the processor/communications ratio for MPI applications shows some promise, although there is too much latency in the TCP/IP based communications for the triangular solutions.

 
Table 7: Empirical Performance Data from the IBM SP1 and SP2 --- EPRI6k --- Complex Variate LU Solver  



David P. Koester
Sun Oct 22 15:31:10 EDT 1995