Compile-time support for efficient data race detection in shared-memory parallel programs
John Mellor‐Crummey · 1993
Data race detection strategies based on software runtime monitoring are notorious for dramatically inflating execution times of shared-memory parallel programs. Without significant reductions in the execution overhead incurred when using these techniques, there is little hope that they will be widely used. A promising approach to this problem is to apply compile-time analysis to identify variable references that need not be monitored at run time because they will never be involved in a data race. In this paper we describe eraser, a data race instrumentation tool that uses aggressive program analysis to prune the number of references to be monitored. To quantify the effectiveness of our analysis techniques, we compare the overhead of race detection with three levels of compile-time analysis ranging from little analysis to aggressive interprocedural analysis for a suite of test programs. For the programs tested, using interprocedural analysis and dependence analysis dramatically reduced ...