Programming, compiling and executing partially-ordered instruction streams on scalable shared-memory multiprocessors

Performance in large-scale shared-memory multiprocessors depends on finding a scalable solution to the memory-latency problem. The author shows that protect consistency (PRC) relaxes previous consistency models with two distinct performance benefits. First, PRC is used to expose and exploit more parallelism in the computation, giving better support to latency tolerance. Second, assuming that visible synchronization directly coordinates changes in the writability of shared data, PRC is used to create more situations where cached data are reusable, giving better support to latency avoidance. The paper evaluates PRC in the context of relaxing intrathread dependences for multithreaded architectures. After the PRC programming notation is described, programming and compiling aspects are examined, and architectural support is discussed.<>