Design and Evaluation of Data Access Prediction Strategies in SDSM Systems

Software Distributed Shared Memory (SDSM) systems provide the shared memory abstraction on top of a message passing hardware, simplifying application programming in these architectures. However, some memory references exhibit long latencies due to remotely cached data. In order to hide this latency, many techniques that propagate data speculatively were developed. This requires that the data access behavior of the applications be determined. Traditionally, many of these techniques were directed at specific data sharing patterns such as producer-consumer and migratory. In this paper, we propose and evaluate generic data access prediction techniques for SDSM systems. By generic we mean that our strategies don't try to detect specific sharing patterns known a priori. The prediction strategies proposed can be divided into two classes: Local Information Predictors (LIP), that are guided only by local information in each processor and Global Information Predictors (GIP) that use the data access pattern of all processors in order to make predictions. Our experimental result show that techniques within both classes can attain high hit ratios in most of the applications evaluated. Overall, the results allow us to conclude that the prediction strategies for data accesses we propose can contribute to increase the performance of current page-based SDSMs significantly.