Extensions of an Empirical Automated Tuning Framework

Abstract

Empirical auto-tuning has been successfully applied to scientific computing applications and web-based cluster servers over the last few years. However, few studies are focused on applying this method on optimizing the performance of database systems. In this thesis, we present a strategy that uses Active Harmony, an empirical automated tuning framework to optimize the throughput of PostgreSQL server by tuning its settings such as memory and buffer sizes. We used Nelder-Mead simplex method as the search engine, and we showed how our strategy performs compared to the hand-tuned and default results. Another part of this thesis focuses on using data from prior runs of auto-tuning. Prior data has been proved to be useful in many cases, such as modeling the search space or finding a good starting point for hill-climbing. We present several methods that were developed to manage the prior data in Active Harmony. Our intention was to provide tuners a complete set of information for their tuning tasks.