Performance analysis of a minimal adaptive router

Abstract

Two classes of adaptive routers, minimal and non-minimal, are emerging as possible replacements for the oblivious routers used in current multicomputer networks. In this paper, we compare the simulated performance of three routers, an oblivious, a minimal, and a non-minimal adaptive router, in a two-dimensional packet switching torus network. The non-minimal adaptive router is shown to give the best performance and the oblivious router the worst. Significantly, however, for many traffic patterns, the minimal adaptive router's performance degrades sharply as the network saturates. Based on an analysis made using several visualization tools, we argue that this performance drop results from nonuniformities introduced for deadlock prevention. Furthermore, this analysis has led us to believe that network balance is an important performance characteristic that has been largely overlooked by designers of adaptive routing algorithms.