An Analysis of Deflection-Based Wormhole Routing with Virtual Channels

We analyze a new wormhole routing scheme, which we call the wormhole deflection with virtual channels (abbreviated WDVC) scheme, that combines wormhole routing with virtual channels and deflection routing to provide efficient lossless communication. We use new analytical models to analyze the performance of the WDVC scheme for the Manhattan Street (MS) network topology. In particular, we examine the effect of the traffic load and the number of virtual channels on the throughput and the length of paths followed by the worms, and compare the analytical results obtained with corresponding simulation results. Our results indicate that wormhole deflections combined with virtual channels is efficient under both light and heavy traffic loads, especially when the number of virtual channels on a link is large. 1 I n t r o d u c t i o n The efficiency in exchanging messages is an important performance parameter of parallel computing systems, and it relies heavily on the switching scheme used for interprocessor communication. Because of its low latency, wormhole routing has evolved as the desired scheme for multiprocessor networks and its performance has been analyzed by several researchers under a variety of models (see, for example, [DaS87], [KID94], [SOD95], [Dua93], and [Dal92]). Conventional wormhole routing allows messages (or worms) to be t ransmit ted as a continuous s t ream of bits occupying multiple links simultaneously. This is accomplished by dividing each message into smaller message units called flits [DaS87], and advancing the flits from the input port to the output port at intermediate nodes, without waiting for the entire message to be received at the node. We distinguish between three types of flits in a message: the header flit used to route the message, the data flits used to carry the message information, and the tail flit used to signify the end of the message. When the header flit arrives at 1 Research supported by ARPA under the MOST project.