Smart DAG Task Scheduling Based on MCTS Method of Multi-strategy Learning

In distributed heterogeneous computing systems, efficient algorithms are crucial for improving system performance. We know that the task scheduling problem based on Directed Acyclic Graph (DAG) has been proven to be an NP-complete problem, making it difficult to find the optimal solution. Although previous research proposed many greedy strategies to solve this problem, these algorithms often have very limited search spaces. To overcome these limitations, we propose a smart DAG task scheduling algorithm based on the Monte Carlo Tree Search (MCTS) method of multi-strategy learning. We design effective state, action, and reward functions to train agents and allow them to adaptively adjust their search strategies. Experimental results prove the effectiveness of our algorithm. Specifically, our algorithm is superior to PSLS, PEFT, and HEFT algorithms in scheduling the maximum completion time under a large number of randomly generated and real-world applications.