Multiprocessor synchronization for concurrent loops

摘要

Execution of concurrent loops on multiprocessor computers often requires synchronizing the processors. Synchronization schemes are surveyed that are suitable for automatic problem decomposition. The model of a shared-memory multiprocessor is used, as is the concurrent-loop paradigm, which is to compile a loop so each processor is assigned a different loop iteration. The discussion covers data dependence, removing synchronization points, random synchronization, pipelining, barrier synchronization, and critical sections.