ANALYZING THE COMBINED EFFECTS OF MEASUREMENT ERROR AND PERTURBATION ERROR ON PERFORMANCE MEASUREMENT

Dynamic performance analysis of executing programs commonly relies on statistical profiling techniques to provide performance measurement results. When a program execution is sampled we learn something about the examined program, but also change, to some extent, the program's interaction with the underlying system and thus its behavior. The amount we learn diminishes (statistically) with each sample taken, while the change we affect with the intrusive sampling risks growing larger. Effectively sampling programs is challenging largely because of the opposing effects of the decreasing sampling error and increasing perturbation error. Achieving the highest overall level of confidence in measurement results requires striking an appropriate balance between the tensions inherent in these two types of errors. Despite the popularity of statistical profiling, published material typically only explains in general qualitative terms the motivation of the systematic sampling rates used. Given the importance of sampling, we argue in favor of the ge neral principle of deliberate sample size selection and have developed and tested a technique for doing so. We present our idea of sample rate selection based on abstract and mathematica