Quantifying WCET reduction of parallel applications by introducing slack time to limit resource contention

In parallel applications, concurrently running tasks cause contention when accessing shared memory. In this paper, we experimentally evaluate how much the Worst-Case Execution Time (WCET) of a parallel application, already mapped and scheduled, can be reduced by the introduction of slack time in the schedule to limit contention. The initial schedule is a time-triggered non-preemptive schedule, that does not try to avoid contention, generated with a heuristic technique. The introduction of slack time is performed using an optimal technique using Integer Linear Programming (ILP), to evaluate how much at best can be gained by the introduction of slack time. Experimental results using synthetic task graphs and a Kalray-like architecture with round-robin bus arbitration show that avoiding contention reduces WCETs, albeit by a small percentage. The highest reductions are observed on applications with the highest memory demand, and when the application is scheduled on the highest number of cores.