Parallel Performance Project Research Paper

Research Paper

Configuration Independent Analysis for Characterizing Shared-Memory Applications
Gheith A. Abandah and Edward S. Davidson
Detailed Technical Report CSE-TR-357-98, University of Michigan, Jan 98.

Abstract

Characterizing shared-memory applications provides insight to design efficient systems, and provides awareness to identify and correct application performance bottlenecks. Configuration dependent analysis is often used to simulate detailed application traces on a particular hardware model. The communication traffic and computation workload generated by the application trace is used as a characterization of this application. This paper demonstrates that configuration independent analysis is a useful tool to characterize shared-memory applications. Configuration independent analysis characterizes inherent application characteristics that do not change from one configuration to another. While configuration dependent analysis is repeated for each target configuration, configuration independent analysis is only performed once. Moreover, configuration independent analysis does not require developing models for the target configurations and is faster than detailed simulation. However, configuration dependent analysis directly provides more information about specific configurations. A combination of the two analysis types constitutes a comprehensive and efficient methodology for characterizing shared-memory applications. In this paper, we show how configuration independent analysis is used to characterize eight aspects of application behavior: general characteristics, working sets, concurrency, communication patterns, communication variation over time, communication slack, communication locality, and sharing behavior. We illustrate the advantages and limitations of this approach by analyzing eight case-study benchmarks from the scientific and commercial domains and interpreting the results.