Just-in-time execution is not always optimal!

Synchronous Data Flow (SDF) is a popular model for analyzing workflows. It is well-known that self-timed scheduling, starting a task when all its prerequisites are satisfied, is throughput-optimal for SDF graphs. Just-in-time scheduling in the presence of deadlines has so far received little attention in the literature. We show that a feasible just-in-time schedule exists if and only if a feasible self-timed schedule exists. Consequently, just-in-time execution is also throughput-optimal. It is, however, not necessarily optimal when tasks may occasionally take longer than anticipated. We experimentally conclude that the likelihood of missing deadlines is significantly higher, 16%, than it is for selftimed scheduling. This shows that Ralph Otten’s way of working, just-in-time scheduling without exception, is not optimal.