A methodology for the design of parallel benchmarks

Hernandez-Gonzalez, Emilio Fernando (1996) A methodology for the design of parallel benchmarks. University of Southampton, Doctoral Thesis.

Abstract

Performance evaluation is necessary to identify the strengths and weaknesses of computer systems and applications. The outcome of performance evaluation is used by computer designers to eliminate bottlenecks in hardware design and improve the balance between the components of computer systems. Since performance evaluation consists in studying computer and program interaction, it is equally useful to improve the efficiency of applications.

This dissertation describes a new methodology for the design of parallel benchmarks. The main objective of the proposed methodology is to bridge the distance between performance evaluation and program design. Three benchmarking techniques are presented in this dissertation and form the basis of the proposed benchmarking methodology; (i) Evaluating Program and Data Transformations, (ii) High-Level Specification of Benchmarks and (iii) White-Box Benchmarking. The first technique evaluates the effect of program change and data transformation on performance. This enables us to design more efficient programs and achieve code re-engineering successfully. The second technique, high-level specification of benchmarks, provides us with a framework to define new benchmarks that evaluate aspects of performance not covered by standard benchmark suites. Finally, white-box benchmarking can be used to obtain performance results that are related to the structure of the benchmark code.

This record has no associated files available for download.

Contributors

Download statistics