Impact of Non-stationary Workload on Resource Reservation Based Slack Reclamation

In open real-time systems with a non-stationary workload, tasks can be dynamically mapped to servers. If a server-based, slack reclamation scheduler is used, each task reserves a fraction of the bandwidth based on its resource estimation. However, the slack time scheduler makes no guarantee to a task beyond its processor capacity reserve, i.e., the scheduler cannot guarantee that a soft real-time task will receive a higher quality of service than is specified by its resource reservation. Nevertheless users often assume that a desired quality of service (QoS) can be achieved using an optimistic resource reservation. Even though existing slack time schedulers perform well in a closed real-time system (with well-behaved applications, stationary workloads, or excessive processor speed), their performance can drop significantly when a task changes its workload. We identify resource underbooking and overbooking problems in non-stationary workloads, resulting in a denial of service (DoS) phenomenon. We solve these problems by preventing DoS attack on the slack time scheduler; we also allow any server to donate slack time even when it is in an idle state. Finally, we introduce a forward donation algorithm that helps existing slack time schedulers reclaim extra slack time. By increasing the robustness of the scheduler, it can be used with broader spectrum of applications in an open real-time system.