Efficient Classification of Application Characteristics by Using Hardware Performance Counters with Data Mining

Abstract

Hardware performance counters in processors are mainly used for low level performance analysis and application tuning by monitoring performance-related hardware events. With the advent of processors with more cores than existing multicore processors and additional high-bandwidth memory, research on the performance analysis of new systems has received increasing attention from the high-performance computing community. Analyzing application characteristics and system features in a new system is essential for computational scientists and engineers who are eager to obtain the best performance of their scientific applications. However, these processors, increased core counts and high-performance resources, make it difficult to understand the correlation between performance-related hardware events. In this paper, we propose a method to simply and quickly classify application characteristics by using a data mining tool without understanding the correlation between hardware events. When we applied the proposed method to NAS Parallel Benchmarks (NPB), the application characteristics were the same as the authorized NPB categories. We show the effectiveness of the proposed scheme in a case study on analyzing the degree of interference between application characteristics.