Parallelism in the Execution of a Routine Knowledge Rule System on the Butterfly (Trademark) Computer

摘要

This project ported a routine knowledge rule system to the BUTTERFLY multiprocessor. The goal was to explore parallelization techniques with an existing rule system originally written for serial execution. The rule system was rewritten to introduce parallelism, and run on a single processor to establish a benchmark for serial operation. The same version, with parallelism enabled, was then run on a 16-node BUTTERFLY multiprocessor. The metering tools on the BUTTERFLY were used to display task behavior and processor utilization. The information gained from these displays was used to guide further experimentation with the granularity of the rules and with the system code to investigate bottlenecks that were lengthening execution time. The project demonstrated that parallelization of routine knowledge rule systems can yield substantial speedup. It also demonstrated that the metering tools on the BUTTERFLY can be used to achieve additional speedup of parallel implementations. The implications of this research are discussed and compared to the findings of research at Carnegie-Mellon University on parallelizing production systems.