Irregular Loop Patterns Compilation on Distributed Shared Memory Multiprocessors
Abstract
The Shared Virtual Memory (SVM) is an interesting layout that handles data storage, retrieval and communication. It affords a shared data programming paradigm on an architecture where the physical memory is actually distributed among several nodes. However, minimizing the false sharing and reducing the synchronization remain important issues for parallel efficiency, either at compile time for regular codes or at runtime for irregular ones. This paper addresses irregular loops compilation on Distributed Memory Parallel Computers (DMPCs) that run an SVM. Solutions are introduced to overcome the poor locality that exhibit irregular loops and a case study on an intel iPSC/2 and a Kendall Square KSR1 is presented.