OpTune: Multi-Point Performance Engineering in Server Systems

Modern server systems encompass multiple components and/or layers containing configuration parameters that can affect performance. Managing the interplay of these parameters is becoming increasingly challenging as the complexity of server systems continue to grow. In this paper, we propose a framework called OpTune to help system administrators tune these parameters. OpTune asks administrators to specify objectives to shape the performance CDFs of systems; e.g., minimize the median response time while keeping the 99th percentile below a target value. In fact, administrators can specify entire target CDFs. OpTune then uses a graphical representation of the system, performance instrumentation and profiling, and manipulations of the profiled CDFs of components to configure the system. We demonstrate the broad utility of OpTune by integrating it into three different, widely-used systems: a Web server, a filesystem emulator, and a MapReduce server cluster. Evaluation results demonstrate that OpTune successfully helps administrators to quickly identify configuration parameter values to best achieve the desired performance behaviors.