Evaluating Two Loop Transformations for Reducing Multiple Writer False Sharing

To simplify the programming of hierarchical and distributed-memory parallel systems, the notion of shared virtual memory (SVM) has been proposed. This abstraction provides the programmer with the illusion of a at global address space and coherence is maintained at the page level. The success of this abstraction depends on the eeciency of page management. This in turn depends on the eeciency of handling false sharing and the resulting ping-pong eeects that it can cause. In this paper, we evaluate two loop transformations for attacking this problem. The rst is a simple, new, compile-time technique for reducing the ping-pong eeects that result from multiple-writer false sharing. The second is our previously-proposed technique for eliminating multiple-writer false sharing itself. These techniques 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.