A Software Support System for Parallel Processing in Multiprocessor Clusters

点击翻译按钮获取中文摘要

This effort is funded as a Research Planning Grant (RPG). The availability of shared-memory, multiprocessor workstation clusters offers the opportunity for the simultaneous application of shared memory and distributed memory parallel processing techniques to a single problem. However, software systems that facilitate this dual-level parallel processing paradigm and allow the programmer, compiler, and run-time system to make the best use of the computational power offered by this platform are not yet available. The long-term goal of this research is threefold: (1) to provide a software system that furnishes advice regarding the partitioning methods and cluster configuration expected to yield the best performance, (2) to develop a communication library that handles both shared and distributed memory communications with a single syntax, and (3) to incorporate both the advisory system and the communication library into a data-parallel compiler that transparently handles interprocessor communication. During the research planning phase, which is being supported by this grant award, four initial steps toward the long-term research goals will be accomplished. (a) The computation and communication characteristics of the multiprocessor cluster will be characterized. (b) Mathematical models of performance of applications that exhibit stencil data dependencies will be developed. (c) The advisory system will be designed and implemented with the initial mathematical models. (d) The current body of research on shared memory and distributed memory communication libraries and data-parallel compilers will be surveyed to determine the best approach for developing the dual-level communication library and incorporating the library into existing compilers. ***