Energy-aware scheduling of conditional task graphs with deadlines on MPSoCs

Abstract

We investigate the problem of scheduling a set of non-pre-emptive tasks with individual deadlines and conditional precedence constraints on MPSoCs (MultiProcessor System-on-Chips) with shared memory such that the total processor energy consumption of all the tasks in each scenario is minimized under two power models, namely the dynamic power model and the total power model, and propose a unified two-phase approach. The approach consists of an offline task scheduler and an online task scheduler. The offline scheduler uses a novel priority scheme to assign each task to a processor, constructs a global schedule, and uses convex NLP (NonLinear Programming) to compute an optimal speed for each task. The online task scheduler dynamically performs task reallocation and task rescheduling, and reassigns a speed to each task to utilize the slack time generated by individual scenarios. We have compared our approach with two state-of-the-art approaches by using 23 benchmarks. The experimental result show that the average improvement and the maximum improvement of our approach over the approach proposed by Ge et al. are 19.2% and 28.6%, respectively, and the average improvement and the maximum improvement over the approach proposed by Malani et al. are 53.2% and 74.2%, respectively.