Configuration independent analysis for characterizing shared-memory applications

Abstract

The paper demonstrates that configuration independent analysis of shared memory applications is useful tool to characterize inherent application characteristics that do not change from one machine configuration to another. Although traditional configuration dependent analysis, or simulation, may directly provide more information about performance on specific configurations, it requires developing a machine model and repeating the analysis for each target configuration. A judicious combination of the two constitutes a comprehensive and efficient methodology. The authors use configuration independent analysis to characterize seven aspects of application behavior: general characteristics; working sets; concurrency; communication patterns; variation over time, and locality; and sharing behavior. Case studies of eight scientific and commercial benchmarks are used to illustrate the advantages and limitations of this approach.