ITR: Optimizing Execution of Parallel Programs on a Cluster of Shared Memory Machines

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

Parallel programs using MPI are widely used in compute-intensive scientific and engineering applications. They perform well on dedicated distributed memory machines and workstation clusters. However, their performance can deteriorate on multiprogrammed shared memory machines (SMMs) or clusters of those machines. This project will optimize execution of parallel programs through both program transformation and efficient run-time support. The resulting programs will deliver robust performance in both dedicated and multiprogrammed SMM clusters. Technically, the work has three aspects: It will study compile-time code transformations to achieve threaded execution of parallel code on a cluster of SMMs, allowing each MPI node to be executed safely as a thread. It will study thread-safe run-time execution and fast lock-free communication that takes advantage of address space sharing among threads within an SMM. It will evaluate and model a variety of scientific applications (including sparse-matrix algorithms with irregular computation, PDE computations with coarse-grain computation, and data-intensive applications) to verify the proposed techniques.