GAT-based Deep Reinforcement Learning Algorithm for Real-time Task Scheduling on Multicore Platform

The scheduling of real-time and dependent tasks on multicore systems plays an important role in influencing system performance. During the scheduling process, multiple constraints should be taken into account, e.g., task dependency, real-time response, and energy efficiency. The complex task scheduling problem makes it difficult to balance solution quality and computation time. Existing works either use time-consuming methods to find optimal solutions or get feasible solutions through heuristic methods. In this paper, we propose a GAT (graph attention network)-based deep reinforcement learning (DRL) algorithm to solve task scheduling problems. This method combines the benefits of deep learning (DL) networks and reinforcement learning (RL) algorithms. It can achieve adaptive learning and adjust information features by constructing the GAT to model the dependencies between tasks. At the same time, we use the soft actor-critic (SAC) algorithm to optimize the task scheduling policy to minimize the makespan of scheduling results. The experimental results show that our method outperforms other scheduling methods regarding scheduling efficiency (the quality of task scheduling problem) and algorithm running time.