Smart DAG Task Scheduling with Efficient Pruning-Based MCTS Method

Abstract

Efficient task scheduling algorithms are important for achieving high performance in heterogeneous computing systems. The DAG (Directed Acyclic Graph) tasks scheduling problem has been extensively studied. However, many previously proposed algorithms are typically heuristic and based on greedy methods, which can not effectively explore large search space of the problem. In this paper, we proposed a smart DAG task scheduling algorithm with efficient pruning-based Monte Carlo Tree Search (MCTS) method. In the task scheduling process, the algorithm adaptively balances the exploration and exploitation of search policies. To further reduce the search space, a new branch-and-bound based pruning method is proposed that significantly improves the algorithm's efficiency. Our experimental evaluation shows the effectiveness of the proposed algorithm. Under a range of different DAGs and heterogeneous configurations, the pruning-based MCTS scheduling algorithm outperforms HEFT, CPOP and PEFT algorithms in terms of scheduling makespan, while restricts the computing overhead below an acceptable threshold.