Overrun handling approaches for overload-prone soft real-time systems

Traditional overrun handling approaches for real-time systems enforce some isolation property at the job or task level. This paper shows that by "relaxing" task isolation, it is possible to efficiently deal with overruns in soft real-time systems with highly variable task execution times and proposes Randomized Dropping (RD), a novel overrun handling mechanism. RD is able to bound task overruns in a probabilistic manner, thus providing "soft" task isolation. The paper shows how to combine RD with priority-driven and rate-based scheduling algorithms, and how to analyze the resulting system. Performance evaluation and comparison between simulation and analytical results are discussed.