An Empirical Study: Mems as a Static Performance Metric

Performance analysis is essential to ensure the non-functional performance requirements of a software system. However, existing runtime-based approaches suffer from the issues of efficiency and platform dependency. In this paper, we investigate the effectiveness of using the mems value to statically estimate the program performance. The mems value, originally proposed by Donald Knuth, is a static and architecture-independent metric used to measure memory access, to estimate program performance statically. We developed an instrumentation tool to record the control flow and measure the mems value by rewriting the source code. Experimental results across ten classical algorithm programs show that execution paths of a program with larger mems values consistently exhibit lower efficiency. Whereas the correlation weakens among different programs. This indicates that the mems maric is best suited for comparing the performance of various paths in the same program.