Programming Shared Virtual Memory on the Intel Paragon Supercomputer

Programming distributed memory systems forces the user to handle the problem of data locality. With message passing the user has not only to map the application to the different processors in a way that optimizes data locality but also has to explicitly program access to remote data. Shared virtual memory (SVM) systems free him from the second task; nevertheless, he is still responsible to optimize data locality by selecting a well-suited work distribution. We describe a programming environment that is based on the Advanced Shared Virtual Memory system, a SVM implementation for the Paragon TM Supercomputer, and on SVM-Fortran, a shared memory parallel programming language with language constructs for work distribution. Programming tools integrate program text and dynamic performance data to help the user in optimizing data locality. Keywords: distributed memory multiprocessors, shared virtual memory, parallel programming tools, programming environments, parallel applications 1 Intro..