Application-Specific Protocols for User-Level Shared Memory

Recent distributed shared memoy (DSM) systems and proposed shared-memory machines have imple-mented some or all of their cache coherence proto-cols in sofiware. One way to exploit the flexibility of this sofiware is to tailor a coherence protocol to match an application’s communication patterns and memoy semantics. This paper presents evidence that this approach can lead to large performance improve-ments. It shows that application-specijic protocols substantially improved the performance of three ap-plication programs—appbt, em3d, and barnes—over carefully tuned transparent shared memo ~ implemen-tations. The speed-ups were obtained on Blizzard, a fine-grained DSM system running on a %?-node Thinking Machines CM-5.