Performance analysis of automatic lookahead generation by control flow graph: some experiments

Abstract

The performance of parallel discrete event models is highly dependent on lookahead, particularly when a conservative algorithm is employed. Unfortunately lookahead is known to be problem-dependent, which restricts the implementations of conservative algorithms. This paper uses a simple queuing network to show how this lookahead affects performance and discusses various techniques for automatic generation of lookahead using control flow graphs (CFGs). These methods are tested on the queuing network simulation running on a CRAY T3E 1200E. Results indicate that the automatic lookahead techniques, though requiring some time to compute, perform as well as the best manually extracted lookahead injected into the parallel program.