A Comparison of Input and Output Driven Routers

Communication in parallel computers requires a low latency router. Once a suitable routing algorithm is selected, an implementation must be designed. Issues such as whether the router should be input or output driven need to be considered. In this paper, we use simulations to compare input driven and output driven routing algorithms. Three algorithms, the Dally-Seitz oblivious router, the *-channels router, and the minimal triplex algorithm are evaluated. Each router is implemented as both an input and an output driven router. Experiments are run for each of the router implementations with seven di erent tra c patterns on both a 256-node two dimensional mesh and torus networks. The results show that in almost all cases, the output driven router matches or outperforms the input driven router. Furthermore, we nd that randomization of output bu er selection in the input driven algorithm increases its performance and substantially reduces the performance discrepancy between the input and output driven algorithms. Although the ndings apply to the routers considered, we believe the results generalize to other routers.