A Response Surface Methodology approach to operating system scheduler tuning

Tuning operating system components is a cyclical process involving setting parameters, evaluating the effect of the settings, making adjustments, and testing again. This is an expensive process, both taking a long time and requiring money to hire people to do it. In this paper we present a statistical approach to tuning of an operating system scheduler using Design of Experiments (DOE) and Response Surface Methodology (RSM). We make use of a benchmark and generate a response surface based on the runtime of the benchmark and three Linux scheduler parameters. We produce a model of the scheduler and optimize the parameter settings, minimizing the number of times the benchmark had to be run to find the optimal settings. In our experiment, we achieved an 11% performance improvement when the Linux scheduler runs the benchmark. We also compared the scalability of the optimized and unoptimized schedulers and discovered that the optimized scheduler does much better in this regard. This was done without prior knowledge of optimal settings for the workload we used.