The Design of a Marker Passing Architecture for Knowledge Processing

Abstract

Knowledge processing is very demanding on com-puter architectures. Knowledge processing generates subcomputation paths at an exponential rate. It is memory intensive and has high communication re-quirements. Marker passing architectures are good candidates to solve knowledge processing problems. In this paper, we justify the design decisions made for the Semantic Network Array Processor (SNAP). Important aspects of SNAP are: the instruction set, markers, relations, propagation rules, interconnec-tion network, and granularity. These features are compared to those in NETL and the Connection Ma-chine. 1 Basic Operations in Knowl-edge Processing The computations that are typical of knowledge pro-cessing require the generation of numerous compu-tation paths that all could potentially be followed in parallel. The process of spawning a number of rela tively independent subcomputations, each of which may spawn other subcomputations, is called bifur-cation. Bifurcation processes appear to be impor-tant for a wide range of knowledge based systems. On a serial computer, the bifurcation of independent subprocesses leads to large computational demands. Even a parallel computer does not have the hard-ware resources to examine all of the parallel paths