On relaxing task isolation in overrun handling to provide probabilistic guarantees to soft real-time tasks with varying execution times

Abstract

Task-level or job-level isolation is commonly used in real-time systems to prevent a job that overruns from affecting jobs belonging to other tasks (task-level isolation) or to the same task (job-level isolation). Although such an isolation provides the desired protection and simplifies analysis, it often results in degraded performance (in terms of deadline meeting) because it limits the exploitation of residual processor time resulting from jobs that underruns. We propose a new overrun handling method called randomized dropping that relaxes the isolation to make effective use of such residual processor time. We apply the proposed randomized dropping to the EDF scheduling and give an analysis technique to compute the probability of deadline meeting. Experimental results show that the proposed overrun handling method outperforms previous approaches based task-level or job-level isolation.