Sampling Issues in the Collection of Performance Data

Many of the performance tools available today sample the performance counters maintained by the system at different rates, with frequencies ranging from once every few seconds to once per minute. This paper is a study of the application of sampling theory to questions such as “What is a good sampling rate to use?”, “Do all metrics need to be sampled at the same rate?”, “What is the overhead of sampling more frequently?”, “How does the sample size affect the accuracy of the collected data?”, “Does system load have an effect on selecting the sampling rate?”, and “How can I tell if the acquired data is suspect?”. We will also review some of the tradeoffs that need to be made among sampling frequency, data accuracy, system overhead and functionality.