Better Real-Time Response for Time-Share Scheduling

Time-share CPU schedulers permeate general-purpose computer systems, yet provide little support for real-time constraints. We demonstrate that by making inferences from the behavior of applications, soft real-time scheduling performance is achievable using a best-effort policy in which the scheduler has no a priori knowledge of application resource needs. In a typical time-share scheduler, recent CPU usage is accounted for in a dynamic scheduling priority, so the behavior pattern of a process implicitly impacts the timeliness of future allocations. To improve the performance of latency-sensitive processes, we developed a CPU scheduler that explicitly uses past behavior to make short-term scheduling decisions, while still preserving the long-term goal of fairness. In this paper, we show that this scheduler, called BeRate, outperforms Linux when scheduling workloads that contain applications with periodic deadlines.