Abstract

Gains due to the parallel execution of Production Systems have been limited by the need for global synchronization before each rule firing. This synchronization can be eliminated by allowing concurrent firing of rules based on the serializability criterion for execution correctness. With this enhancement however, the execution demands of the local Rete network within each processor again resurfaces as the performance bottleneck. This paper presents a novel organization that allows the use of multiple functional units to execute the Rete subnetwork within each processor. This organization is reminiscent of current-generation superscalar architectures for numerical processing and indeed borrows some of their synchronizing mechanisms. We present results from a detailed event-driven simulator to quantify the benefits of employing multifunctional units, and show that using a modest number of processors, each containing a small number of functional units, provides the most cost-effective sol...

Citations