Experiments with a Knowledge-Based System on a Multiprocessor

This paper documents the results we obtained and the lessons we learned in the design, implementation, and execution of a simulated real-time application on a simulated parallel processor. Specifically, our parallel program ran 100 times faster on a 100-processor multiprocessor. The machine architecture is a distributed-memory multiprocessor. The target machine consists of 10 to 1000 processors, but because of simulator limitations, we ran simulations of machines consisting of 1 to 100 processors. Each processor is a computer with its own local memory, executing an independent instruction stream. There is no global shared memory; all processes communicate by message passing. The target programming environment, called Lamina, encourages a programming style that stresses performance gains through problem decomposition, allowing many processors to be brought to bear on a problem. THe key is to distribute the processing load over replicated objects, and to incresase throughput by building pipelined sequences of objects that handle stages of problem solving. We focused on a knowledge-based application that simulates real-time understanding of radar tracks, called Airtrac. This paper describes a portion of the Airtrac application implemented in Lamina and a set of experiments that we performed. We confirmed the following hypotheses: 1) Performance of our concurrent program improves with additional processors, and thereby attains a significant level of speedup. 2) Correctness of our concurrent program can be maintained despite a high degree of problem decomposition and highly overloaded input data conditions.