Evaluating Two Loop Transformations for Reducing Multiple Writer False Sharing

Shared virtual memory (SVM) simplifies the programming of parallel systems with memory hierarchies and physically distributed address spaces, by providing the illusion of a flat global address space where coherency is maintained at the page level. The success of the SVM abstraction depends on efficient page management, which in turn depends on the efficient handling of false sharing and the resulting ping-pong effects that it can cause. We evaluate two loop transformations for attacking this problem. The first is a simple, new technique for reducing the ping-pong effects that result from multiple-writer false sharing. The second is our previously-proposed technique for eliminating multiple-writer false sharing itself. Both have been implemented in the Fortran-S compiler, which generates code that runs on the iPSC/2 under the KOAN SVM. Preliminary performance results are presented.