Automating the Coordination of Interprocessor Communication
Jingke Li, Marina Chen · 1990
This paper presents methods for ensuring correct synchronization and scheduling of message-passing in the context of compiling shared-memory programs onto distributed-memory machines. We show that from a given source loop nest, there corresponds a maximum granularity where the computation can go on without the need for any communication, and a communication window within which a communication command must occur and can occur anywhere legally. Better overall efficiency can then be achieved by playing with the granularity parameter using more frequent communication than that for the maximum granularity case.