Quantitative Performance Evaluation for Program Optimization by Profiling

摘要：

In order to explore maximum computation capacity, optimization is often necessary for programs that run on some specified hardware platform. Performance evaluation and analysis are thus fundamental to understand where the performance bottlenecks are and how they could be eliminated. The traditional qualitative evaluation and analysis only indicate where the benefits are from, but do not quantitatively measure how much performance improvement the optimization could achieve. In this paper, we present an analytical model, together with the data collected by a performance profiler, to quantitatively evaluatethe performance of a program. Our model focuses on memory hierarchy performance, upon which the program performance of modern architecture is significantly dependent. The model is validated via a series ofexperiments analyzing four different implementations of the matrix multiplication algorithm, which exhibit different memory access patterns.The experimental results show that the estimated execution time is reasonably close to actual execution time. Thus, our model could be used to guide the process of program optimization and measure the achievement of optimization.

关键词： Program Optimization Performance Model Quantitative performance Evaluation Memory Hierarchy Performance Profiler

作者: Qiankun Miao Yunquan Zhang Guangzhong Sun Guoliang Chen

作者单位: Department of Computer Science,University of Science and Technology of China,Hefei,230027,P.R.China; Laboratory of Parallel Computing,Institute of Software,Chinese Academy of Sciences,Beijing 100190,P. Department of Computer Science,University of Science and Technology of China,Hefei,230027,P.R.China

会议类型: 国际会议

会议名称: The Inaugural Symposium on Parallel Algorithms, Architectures and Programming(并行算法、结构和编程国际研讨会)

会议地点: 广州

会议语种:英文

页码: 144-162

在线出版日期: 2008-09-16（万方平台首次上网日期，不代表论文的发表时间）