Optimal topology for distributed shared-memory multiprocessors: hypercubes again?
|Type of Publication:||In Proceedings||Keywords:||distributed memory systems;hypercube networks;memory protocols;message passing;performance evaluation;shared memory systems;|
|Pages:||205 - 12|
optimal topology;distributed shared memory multiprocessors;DSM;direct interconnection network;message traffic;multicomputers;higher dimensional topologies;low dimensional topologies;2D mesh;hypercube topologies;router model;network load;cache coherence protocols;average latency;
Many distributed shared memory multiprocessors (DSM) use a direct interconnection network to implement a cache coherence protocol. An interesting characteristic of the message traffic produced by coherence protocols is that all the messages are very short. Most current multicomputers use low dimensional meshes or tori because these topologies usually achieve a higher performance. However, when messages are very short, latency is mainly dominated by the distance traveled in the network. As a consequence, higher dimensional topologies may achieve a lower latency than low dimensional topologies. We compare the 2D mesh and the hypercube topologies assuming a very detailed router model. Network load has been modeled taking into account the traffic produced by cache coherence protocols. Performance results show that average latency for hypercubes is slightly lower than for meshes. Moreover, hypercubes achieve a much higher throughput than meshes, making them suitable for DSMs