Measuring program similarity for efficient benchmarking and performance analysis of computer systems

Abstract

Computer benchmarking involves running a set of benchmark programs to measure performance of a computer system. Modern benchmarks are developed from real applications. Applications are becoming complex and hence modern benchmarks run for a very long time. These benchmarks are also used for performance evaluation in the early design phase of microprocessors. Due to the size of benchmarks and increase in complexity of microprocessor design, the effort required for performance evaluation has increased significantly. This dissertation proposes methodologies to reduce the effort of benchmarking and performance evaluation of computer systems. Identifying a set of programs that can be used in the process of benchmarking can be very challenging. A solution to this problem can start by identifying similarity between programs to capture the diversity in their behavior before they can be considered for benchmarking. The aim of this methodology is to identify redundancy in the set of benchmarks and find a subset of representative benchmarks with the least possible loss of information. This dissertation proposes the use of program characteristics which capture the performance behavior of programs and identifies representative benchmarks applicable over a wide range of system configurations. The use of benchmark subsetting has not been restricted to academic research. Recently, the SPEC CPU subcommittee used the information derived from measuring similarity based on program behavior characteristics between different benchmark candidates as one of the criteria for selecting the SPEC CPU2006 benchmarks. The information of similarity between programs can also be used to predict performance of an application when it is difficult to port the application on different platforms. This is a common problem when a customer wants to buy the best computer system for his application. Performance of a customer's application on a particular system can be predicted using the performance scores of the standard benchmarks on that system and the similarity information between the application and the benchmarks. Similarity between programs is quantified by the distance between them in the space of the measured characteristics, and is appropriately used to predict performance of a new application using the performance scores of its neighbors in the workload space.