An Empirical Comparison of Networks and Routing Strategies for Parallel.pdfVIP

Figure 4: Comparison of packet routing latency under constraint of equal interconnect area. Figure 5: Comparison of wormhole routing latency under constraint of equal pin-out. Figure 2: Comparison of packet routing latency under constraint of equal bisection width. Figure 3: Comparison of packet routing latency under constraint of equal pin-out. can be dominant [11]. Maximum throughput can be readfrom the latency graphs by looking for the load rate atwhich the network saturates.Figures 2 through 4 show packet routing simulationresults under constant bisection width, constant pin-out,and constant area constraints, respectively. Simulationsare shown for several values of n up to n = 4096. Messagelengths of 320 bits are used throughout.The most striking aspect of the packet-routing sim-ulation results in Figures 2 through 4 is that the meshalways performs very well in comparison to the other net-works despite the use of the unit wire delay model. Whilethe best low-load latency is obtained with the fat-tree un-der constant bisection and constant pin-out constraints(for large networks), it is surprising that the performanceof the fat-tree is not generally better than what is shownby our simulations, particularly under the sort of areaconstraint that motivated study of the fat-tree. Perfor-mance of the fat-tree and fat-pyramid might be betterwith the more area-ecient variation in [8, Secs. II{III].Also interesting is that for the most part, the packet rout-ing graphs look qualitatively very similar to those ob-tained from wormhole routing with a reasonable rangeof worm lengths. (As would be expected, however, thepacket routing results tend to show higher average laten-cies and higher maximum throughput.) Only in the caseof constant pin-out did the choice of packet routing ver-sus wormhole routing cause some change in the rankingof networks by low-load latency; our wormhole routingresults for constant pin-out are shown in Figure 5.References[1] S. Abraham and K. Padmanabh