基于并行Benchmark的高性能机实用测试与评价方法

A Practical Method Based on Parallel Benchmarks for Measuring and Evaluating High-Performance Parallel Computers

文中分析了传统基于并行基准测试程序测试方法的不足,针对目前高性能计算机处理机台数多的特点,给出了实用的测试方法,推广了传统加速比的概念,提出了一种可扩展性评价方法。使用文中给出的方法,可方便地对高性能计算机进行测试和性能评价,同时可以使用小规模的并行机测试结果预测大规模并行机的性能。最后给出了NPB在某高性能计算机上的测试和性能评价结果。

In this paper, we discuss the drawbacks of the traditional method based on parallel benchmarks for measuring and evaluating high-performance parallel computers. According to the characteristics that high-perfor-mance parallel computers have large numbers of processors, we give a practical measuring method for them. In order to evaluate the performance of a supercomputer, we generalize the traditional speedup in accordance with our method. Scalability has become very important for supercomputers, but we think that the old scalability metrics are not practical, so we give a new metric which can give the scalability very conveniently using the data from our new measuring method. Finally, we give a few examples to show how to measure and evaluate high-performance parallel computers using our new method.