ALGORITHMS FOR DYNAMIC SOFTWARE CACHE COHERENCE

In this paper, we investigate a class of coherence strategies in which an abstraction for shared data at the program-level, referred to as S hared Regions (SR), is used to manage caches dynamically through softw are. The practical value of these strategies is measured by their perf ormance relative to existing hardware coherence protocols, and the com plexity of the SR programming interface. We present detailed quantitat ive results highlighting the performance of a wide array of SR coheren ce algorithms, including some novel algorithms introduced in this pape r that use direct cache-to-cache data transfers via software to improv e performance. These algorithms are studied using execution-driven sim ulation and compared to a representative hardware strategy for a suite of parallel applications. The experimental results show that the best SR coherence strategy for each application is comparable to or signif icantly better than the representative hardware strategy for all of th e applications that we examine. Our study of programming complexity in SR finds that, for the types of applications that we study, inserting SR annotations is a relatively simple and methodical task. (C) 1995 A cademic Press, Inc.