Simulating Virtual Cut-through and Wormhole Routing in a Clustered Torus

Communication with uniformly distributed destinations, with locality and with hot-spots are simulated in a 2D clustered torus using virtual cutthrough and wormhole routing. Assuming uniform distributed destinations wormhole routing has a better performance than virtual cut-through when both have the same buffer size; wormhole routing with a buffer size of one flit has a worse performance. Partial virtual cut-through has a better performance than full virtual cut-through. With uniform distributed destinations a 3-clustered torus network has a better maximum throughput per processor than a 4-clustered torus network. It is concluded that the more locality there is, the lower is the latency and the better is the throughput. If the probability a message is sent to a hot-spot decreases or messages are sent to more than a single hot-spot, the throughput increases, the latency decreases and the average distance a delivered message has traveled over is higher. Keywords parallel co..