Energy-Aware Scheduling of Periodic Conditional Task Graphs on MPSoCs
Umair Ullah Tariq, Hui Wu · 2017
We investigate the problem of scheduling a set of periodic conditional task graphs of non-preemtible tasks on a MPSoCs (Multi Processor System-on-Chip) with shared memory such that the total expected processor energy consumption of the tasks in each scenario is minimized under two power models, namely dynamic and static power model, and propose a novel offline scheduling approach. Our approach consists of a novel two-phase task scheduler that aims at minimizing total worst-case utilization of each processor and an optimal task execution speed selection algorithm using convex NLP (Non-Linear Programming). Furthermore, we propose an O(1) time online DVS (Dynamic Voltage Scaling) heuristic that assigns each task a speed online. Our experimental results show that our two-phase scheduler achieves a 95.2% success rate of constructing a feasible schedule, compared to a 42% success rate of the state-of-the-art. For energy saving, our offline scheduling approach achieves an average improvement, a maximum improvement and a minimum improvement of 8.03%, 14.08%, and 4.8%, respectively over our online DVS heuristic.