Schemes for reducing communication latency in regular computations on DSM multiprocessors

This paper proposes two schemes for reducing the communication latency in the computations with local and regular communication patterns. The primary scheme is the local protocol for enhancing spatial locality of the coherence actions, and the second is for the communication via stream data. The local protocol is a subsidiary of a traditional coherence protocol and activated on a memory block basis. If activated on a memory block, the domain of coherence actions is confined into the caches that are currently sharing that block. Otherwise, the domain involves the memory as well as those caches. It is easy to efficiently communicate via stream data in the local coherence domain. Our experimental results show that one of the three proposed local protocols, the tampering protocol, significantly improves the performance as compared with the conventional protocol. The communication via stream data contributes to a further performance improvement