Impact of Predictive Switching in 2-D Torus Networks
Tsutomu Yoshinaga, Hirokazu Murakami, Michihiro Koibuchi · 2007
Predictive switching is a technique for reducing message latency in parallel computer networks. It tries to decide traversal paths of messages by utilizing a prediction mechanism so that processing time for message headers can be shortened. A key issue of predictive switching is the overhead of prediction failures. This paper presents simple and efficient treatments of prediction failures. Our proposal includes three schemes. The first scheme is arranging predictive and non-predictive routers in a network to safely detect and discard mis-predicted packets. The second is additional hardware to reduce occurrences of mis-predicted packets. The third scheme is to shorten the mis-predicted packets. We show the impact of predictive switching embodying the three schemes for k-ary 2-cubes (k = 8, 16, 32) with dimension- order routing. Our simulation results demonstrate that we can reduce average message latency by minimizing the prediction-failure overhead. Network saturation throughput is also improved when the predictor's accuracy is high.