Reference: Muliawan, D. Performance Evaluation of a Parallel Knowledge-Based System. KSL, June, 1989.
Abstract: This paper discusses the quantitative and qualitative performance of a module of a parallel knowledge-based system for tracking and classifying aircraft, called Airtrac. The Airtrac system is built to gain some understanding of the potential speed up through concurrency of reasonably large and complex continuous signal understanding systems. Airtrac runs on CARE, a simulated distributed-memory, asynchronous message-passing multicomputer. Evaluating the performance of a continuous parallel knowledge-based system such as Airtrac is difficult. The simple approach of timing its execution would not work, since the system is continuous. Furthermore, the performance is usually multidimensional and cannot easily be expressed into a single number. The notions of latency, excess ratio, sustainable data rate, and capacity are instead used to rate the performance of the system. The paper reports the effects of important high-level control strategies on the system performance, the effects of varying the frequency and width of the input data across different numbers of processors, and some possible speed up curves of the overall system performance as a function of the number of processors. Finally conclusions are presented in the relationship between the quantitative and qualitative performance of Airtrac, and in the potential speedup of large and complex parallel knowledge-based systems.