Towards Automatic Performance Tuning

Abstract

When the computing environment becomes heterogeneous and applications become modular with reusable components, automatic performance tuning is needed for these applications to run well in different environments. We present the Active Harmony automated runtime tuning system and describe the interface used by programs to make applications tunable. We present the optimization algorithm used to adjust application parameters and the Library Specification Layer which helps program library developers expose multiple variations of the same API using different algorithms. By comparing the experience stored in a database, the tuning server is able to find appropriate configurations more rapidly. Utilizing historical data together with a mechanism that estimates performance speeds up the tuning process. To avoid performance oscillations during the initial phase of the tuning process, we use improved search refinement techniques that use configurations equally spaced throughout the performance search space to make the tuning process smoother. We also introduce a parameter prioritizing tool to focus on those performance critical parameters. We demonstrate how to reduce the time when tuning a large system with many tunable parameters. The search space can be reduced by checking the relations among parameters to avoid unnecessary search. In addition, for homogeneous processing nodes, we demonstrate how to use one set of the parameters and replicate the values to the remaining processing nodes. For environments where parameters can be divided into independent groups, an individual tuning server is used for each group. An algorithm is given to automatically adjust the structure of cluster-based web systems and it improves the system throughput up to 70%. We successfully apply the Active Harmony system to a cluster-based web service system and scientific programs. By tuning the parameters, Active Harmony helps the system adapt to different workloads and improve the performance up to 16%. The performance improvement cannot easily be achieved by tuning individual components for such a system and there is no single configuration that performs well for all kinds of workloads. All the design and experimental results show that Active Harmony is a feasible and useful tool in performance tuning.