Some guidelines for proportional share CPU scheduling in general-purpose operating systems

Our premise is that since there already exists a large, mature body of literature on real-time scheduling in general-purpose operating systems, it is time to spend more effort deciding which of these algorithms should be used and when, and less effort on generating new algorithms. In this paper we focus on proportional share schedulers. We introduce the notion of pessimism-the proportion of overreservation required for an application to meet real-time deadlines when scheduled by proportional share schedulers that have bounded allocation error. We study the implications of pessimism and its effect on the selection of scheduling algorithm and scheduling quantum size, and also the interaction of quantum size and context switch overhead. Finally, we examine the implications of these tradeoffs for the design of applications and schedulers.