作者: Harjinder S. Sandhu , Benjamin Gamsa , Songnian Zhou
关键词:
摘要: The effective management of caches is critical to the performance applications on shared-memory multiprocessors. In this paper, we discuss a technique for software cache coherence tht based upon integration program-level abstraction shared data with . abstraction, called Shared Regions, explicitly relates synchronization objects they protect. Cache algorithms are presented which use information provided by region primitives, and ensure that regions always cacheable processors accessing them. Measurements experiments Region approach multiprocessors multiprocessor shown. Comparisons other strategies, including user-controlled strategy an operating system-based strategy, show able deliver better performance, relatively low corresponding overhead only small increase in programming effort. Compared compiler-based Regions still performs than compiler can achieve 90% accuracy allowing cacheing, as long few hundred bytes or larger, re-used times cache.