Skip to main content
U.S. Department of Energy
Office of Scientific and Technical Information

Instrumenting distributed memory parallel systems

Book ·
OSTI ID:5124058
 [1]
  1. Illinois Univ., Urbana, IL (United States). Dept. of Computer Science

In parallel architectures, distributed memory systems consist of a set of interacting nodes. Each node contains a processor with some locally addressable memory, a communication controller capable of routing messages without delaying the processor, and a small number of connections to other nodes. Because the nodes do not share memory, the cooperating tasks of a parallel algorithm must execute asynchronously on different nodes and communicate solely via message passing. As charged by the organizers of the workshop on Instrumentation for Future Parallel Systems, the working group for distributed memory instrumentation consisered tools and techniques for measuring the performance of distributed memory parallel systems. This paper summarizes the author's impressions of the working group consensus namely, that software instrumentation of node operating systems to capture communication characteristics is needed in the short term. In the longer term, hardware support is needed for capture of event traces with global, synchronous time stamps.

OSTI ID:
5124058
Country of Publication:
United States
Language:
English