Swarm–Intelligence-Based Task Scheduling for Reliability Optimization of Integrated CPU–GPU Edge Platforms in Cyber–Physical–Social Systems

With the increasing demand for low power and high performance, central processing unit–graphic processing unit (CPU–GPU) heterogeneous multiprocessor systems-on-chip (MPSoCs) are widely used in edge computing-supported cyber–physical–social systems (CPSSs). However, few existing task scheduling approaches proposed for CPU–GPU heterogeneous MPSoCs cope with reliability issues, especially when considering the cross-layer propagation of soft errors. This article focuses on improving the soft error reliability (SER) of executing independent real-time applications on an integrated CPU–GPU edge platform in CPSSs under deadline and power constraints. To this end, we first formulate the makespan and power-aware SER optimization problem, which considers the cross-layer propagation of soft errors. To solve the optimization problem, we propose a three-stage swarm intelligence-based task scheduling scheme that decides the number of replicas for tasks, the assignment of tasks and replicas to CPU/GPU cores, and the operating frequency of tasks and replicas executed on the cores. Finally, we validate our proposed task scheduling scheme through extensive experiments on a simulated MPSoC and two NVIDIA CPU–GPU heterogeneous edge platforms. Evaluation results show that our scheme is superior to two state-of-the-art algorithms in many aspects, with an 8.80% increase in reliability and a 13.67% reduction in makespan.