An Associative Memory Architecture for Concurrent Production Systems 1. Motivation

This paper presents a novel parallel architecture for production systems. It relies on modern asso-ciative memory techniques to construct an environment in which productions can be red in parallel even before the actions of previously red productions are fully processed. This approach is made possible by the use of the serializabil-ity criterion for correctness. An approximate solution for the rule partitioning problem is presented. Measurements of speedup obtained from a detailed event-driven simulator indicate the potential of this new architecture. The interest in new research towards acceleration of knowledge systems is sustained by an extensive and prootable market that is in need of handling massive databases at ever increasing speeds 15, 6]. Kuo and Moldovan 9] and Amaral and Ghosh 2] survey various approaches to improve Production Systems (PS) execution speed. In spite of the increasing popularity of parallel architectures, the architectures providing best cost/performance are compiled Production Sys-Sponsored in part by Conselho Nacional de Desen-volvimento Cient co e Tecnoll ogico (CNPq) and Pon-tif cia Universidade Catt olica do Rio Grande do Sul (PU-CRS)-Brazil. tems (PS) running on general purpose unipro-cessors 11]. The search for an eeective parallel PS machine follows two schools of thought. The rst believes that little improvement can be obtained in the performance of parallel PS without the introduction of signiicant changes to PS language semantics 14]. The second tries to maintain the most attractive aspect of production systems, that is, the simplicity of \OPS5-like" semantics 3], while seeking better performance through architectural improvements 10, 13]. The original OPS5 implements the commuta-tivity criterion of correctness, which demands that a parallel implementation produces the same results as any sequential execution of productions 7]. The selection strategy of OPS5 allows only the most recent and speciic production to re at any cycle. Programmers often rely on this strategy to ensure correctness. The architecture presented in this paper maintains most of the OPS5 semantics; it modiies, however, the selection strategy and uses serializability as a cor-rectness criterion 12, 13]. The serializability criterion allows any matched production to re at any time as long as at least one sequential execution of the productions produces the same results as the parallel one, Serializability is less restrictive and allows more parallelism in the execution of a production system. The drawback is that