Strategies optimization and integration in DSM

Abstract

In large-scale parallel computing that may contain many nodes, a computing task is often divided into several sub-tasks running on a large platform. We present a group-based DSM model named GDSM. In our solution, sub-tasks are to be mapped to groups, which applies different techniques based on the distinct characteristics of inter-group and intra-group data sharing. A series of alternatives are employed with respect to the intra-group and inter-group property of this mechanism: (1) Consistency model: Release Consistency vs. Scope Consistency; (2) Coherence protocol: multiple-writer and write-update protocol vs. single-writer and write-invalidate protocol; (3) Granularity: fine-grain vs. coarse-grain. Our strategy to combine grouping with consistency and granularity switch enhances flexibility, cuts down the overhead of consistency maintenance, increases the parallelism between groups and thus promotes system performance. We also put forward the architecture of Protocol Engine based on group implementing the GDSM scheme, the engine that achieves message-forwarding transparency, clear hierarchical structure and encapsulation of group.