A Deep Reinforcement Learning Scheduling Algorithm for Heterogeneous Tasks on Heterogeneous Multi-Core Processors

Heterogeneous multi-core processor systems are common complex processing environments for scheduling DAG application tasks. Deep reinforcement learning has become a popular solution for scheduling in these systems due to its superior direct perception decision-making and handling of high-dimensional state-action spaces. This paper investigates a scheduling problem in a heterogeneous multi-core processor environment. First, a transformer-based graph convolutional neural network extracts and encodes system environment information. The decision space is then reduced by splitting task selection and processor allocation, where the former is managed by a deep neural network learning to select nodes, and the latter uses a heuristic scheduling algorithm. The entire scheduling process is framed as a Markov decision problem. Therefore, the PPO algorithm with dynamic adjustment of the clipping factor, combined with the advantage actor-critic network, is used to train, optimize, and evaluate the algorithm, seeking the optimal scheduling strategy. The training process used a reward function for the time required to complete task scheduling, and experiments were conducted in different environments. The results showed that compared with other algorithms, the algorithm proposed in this paper reduced the task scheduling time overhead of heterogeneous multi-core processor systems by 10.89%.