Establishing Causality as a Desideratum for Memory Models and Transformations of Parallel Programs

In this paper, we establish a notion of causality that should be used as a desideratum for memory models and code transformations of parallel programs. We introduce a Causal Acyclic Consistency (CAC) model which is weak enough to allow various useful code transformations, yet still strong enough to prevent any execution that exhibits “causal cycles” that may be caused by the Java Memory Model (JMM) [18]. For memory models, we introduce a graph model called causality graph that can be used to analyze if a particular program execution violates causality. By using causality graph, we show that a popular memory model (such as the Java memory model) can lead to program executions that exhibit causality violations with respect to our notion of causality. For code transformations, we establish criteria to identify transformations that are causality-preserving which do not result in any execution that exhibits causality violation. We showed that the CAC model allows all the causality-preserving transformations. Finally, we present preliminary experimental results for a load elimination optimization to motivate the performance benefit of using the CAC model relative to the Sequential Consistency (SC) model which is the most basic memory model. For the benchmark program studied, the number of getfield operations performed was reduced by 37.9% by using the CAC model instead of the SC model, and the execution time on a 16-core processor was reduced by 46.2%.