Proportional-Share Scheduling: Implementation and Evaluation in a Widely-Deployed Operating System

This paper explores the feasibility of using lottery scheduling , a proportional-share resource management algorithm, to schedule processes under the FreeBSD operating system. Proportional-share scheduling enables exible control over relative process execution rates and processor load insulation among groups of processes. We show that a straight implementation of lottery scheduling performs worse than the standard FreeBSD scheduler. This initial result prompted us to extend lottery scheduling. Except for one test we run, our resulting system performs within one percent of the FreeBSD scheduler. We describe our design, evaluate our implementation, and relate our experience in deploying our lottery scheduler on production machines. 1 Introduction This paper explores the feasibility of proportional-share resource management in a modern and well-understood operating system. Specically, we employ lottery scheduling Waldspurger & Weihl 1994] to allocate processor time in FreeBSD 2.2.5R. We describe and evaluate our implementation , a hybrid of the FreeBSD scheduler and lottery scheduling, and relate our experiences in using this sched-uler on production machines. To begin, we motivate this paper with a summary of the beneets of proportional-share resource management. Running processes make progress by consuming system resources such as processor cycles, network bandwidth, and storage. Considering processor cycles in particular, mul-tiprogrammed systems allocate xed quanta of processor time to running processes. A CPU scheduling algorithm based on proportional-share scheduling enables exible control over the relative rates processes are allocated processor time quanta. For CPU-bound workloads, such as those found in some scientiic and engineering environments, these rates are equal to the relative rates processes consume CPU. For example, consider two independent CPU-intensive scientiic simulations. These simulations begin running at timestep 0 and output intermediate results at the end of every timestep to be used by a 3-D rendering application. A researcher wants to visualize the results of these timesteps synchronously , but unfortunately the rst simulation requires less computation and thus runs three times faster. Using a proportional-share process scheduler, the researcher can easily adjust the relative execution rates of the simulations so that the second process is chosen to run three times as often.