A new injection limitation mechanism for wormhole networks

Abstract

Wormhole switching has been widely applied to the interconnection networks of parallel systems as well as System Area Networks, and Local Area Networks, largely because of its efficiency and performance merits. Examples include the Myrinet of Myricom Inc as well as most of the newly developed parallel systems. True Fully Adaptive Routing (TFAR) Algorithms have demonstrated their suitability for wormhole switched networks due to their unrestricted Adaptivity and moderate resource requirements. Wormhole switching has proven to be the most popular switching technique targeted for interconnection networks of message-passing multicomputers as well as SANs, and LANs. TFAR Algorithms have also been gaining favor for application in wormhole switched networks due to their highly adaptive and moderate hardware requirements. Wormhole switched networks have associated drawbacks however, as they generally suffer from performance degradation beyond the saturation point due to channel congestion. Fully adaptive algorithms are vulnerable to cyclic dependencies, which are precursors to deadlock formations. Consequently the frequent occurrence of deadlocks can further degrade the performance and stability characteristics of these networks. Injection limitation techniques were recently introduced in an attempt to countermeasure these drawbacks and effectively contain their impact on the performance of the network. This paper proposes a new injection limitation mechanism and its performance evaluation. The new mechanism is named Congestion Level Injection Control (CLIC). This mechanism attempts to provide a solution for these problems and improve the overall performance of the network. The new mechanism is centered on congestion level estimation in the network using only local information at each node. The mechanism subsequently prevents the injection of new packets if the network is deemed to be highly congested or possibly close to its saturation point. The performance of the CLIC mechanism has been compared with other competing schemes. Our results have shown that CLIC has superior performance when compared to other competing schemes.