Prefetching strategies for partitioned shared objects

The author has developed a parallel programming model in which partitioned shared objects encapsulate the state and the distribution of data structures in shared-data objects. In this model, the state of an object can be accessed through sequential or parallel operations. This paper describes the object model and partition dependency graphs (PDGs), a mechanism our runtime system uses to prefetch data, reduce access overhead, and increase the overlap between computation and communication during the execution of parallel operations. This mechanism can be applied to statically as well as dynamically allocated objects, i.e., objects for which size and distribution may or may not be known at compile time. This paper illustrates the effectiveness of our prefetching mechanism with two applications: successive overrelaxation (SOR) and fast Fourier transform (FFT). With prefetching, we obtained, respectively, up to 30% and 10% performance improvement