Experience with Using a Performance Predictor during Development: A Distributed Storage System Tale

Developing a distributed system is a complex and error-prone task. Properly handling the interaction of a potentially large number of distributed components while keeping resource usage low and performance high is challenging. The state-of-the-practice on performance evaluation focuses on employing profilers to detect and fix potential performance problems, but lacks decision support information about when profiling effort should stop. To help address this issue, we propose the use of a prediction tool to estimate the expected performance of a complex system, and describe our experience with employing this tool to support the development of a distributed storage system.