Experimental Benchmarks and Initial Evaluation of the Performance of the PASM System Prototype

The work reported here represents experiences with the PASM parallel processing system prototype during its first operational year. Most of the experiments were performed by students in the Fall semester of 1987. The first programming, and the first timing measurements, were made during the summer of 1987 by Sam Fineberg. The goal of the collection of experiments presented here was to undertake an Application-driven Architecture Study of the PASM system as a paradigm for parallel architecture evaluation in general. PASM was an excellent vehicle for experimenting with this evaluation technique due to its unique architectural features. Among these are: 1. A reconfigurable, partitionable multistage circuit-switched network. 2. Support for both SIMD and MIMD programs. 3. Ability to execute hybrid SIMD/MIMD programs. 4. An instruction queue which allows overlap of control-flow and data manipulation between micro-control (MC) units and processing elements (PE). It had been hypothesized that superlinear speed-up over the number of PEs could be attained with this feature, and experimental results verified this. 5. Support for barrier synchronization of MIMD tasks. This feature was exploited in some non-standard ways to show the ability to decouple variant length SIMD instructions into multiple MIMD streams for an overall performance benefit. This type of study is expected to continue in the future on PASM and other parallel machines at Purdue. This report should serve as a guide for this future work as well