Ordinal Optimization of Communication Network Performance: A New Look Based on use of the Connection Machine

Abstract : Ordinal optimization can be an effective technique for efficiently finding nearly optimal solutions to complex problems. The motivation behind this approach is that finding the optimal solution (or control policy) is often too costly or time consuming, although a suboptimal solution may provide sufficiently good performance. In earlier studies on sequential machines, we demonstrated the effectiveness of ordinal optimization based on the Standard Clock (SC) parallel simulation technique. In this report we study the use of SC and ordinal optimization techniques on the massively parallel Connection Machine CM-5E. The use of the CM-5E has greatly extended the size of problems that can be addressed. For example, whereas our studies on sequential machines were typically limited to wireless networking examples with up to 8 transceivers per node, the use of the CM-5E has permitted the study of examples with up to 4,000 transceivers per node, thus permitting the study of examples with the dimensions of high-speed networks. We address self-regulation and scalability properties of the solutions, as well as the determination of good solutions for large, finely-quantized systems. The CM-5E has enabled us to demonstrate that good solutions can be found quickly, and often without the need for high-performance computer resources.