Timely Recovery from Task Failures in Non-preemptive, Deadline-driven Schedulers
Michael Short, Imran Sheikh · 2010
Although preemptive scheduling mostly dominates non-preemptive scheduling from a feasibility perspective, developers of systems with resource constraints may sometimes choose to implement the latter. Amongst the available techniques for scheduling these systems, non-preemptive EDF (npEDF) is known to be an attractive option. However as with most non-preemptive forms of scheduling, problems may still arise due to the single-tasking nature of its operation. In particular npEDF can be highly susceptible to complete system failures (`timeline breaks') due to errors affecting only a single task. This paper will present a simple Overrun Detection and Recovery Mechanism (ODRM) that may help to alleviate this problem, by detecting task failures in such a fashion that subsequent task deadlines are not missed in a `domino-style' manner. It also allows for the optional execution of a recovery handler. The technique is applied to a case study consisting of a real-time control system for an unstable process; the paper describes initial results which indicate ODRM allows for an improved ability to tolerate task failures, and has a minimal impact on scheduling overhead.