HeterSched: Structure-Aware Reinforcement Learning for Heterogeneous Distributed Task Scheduling

Abstract

To address the scheduling challenges posed by diverse task types, significant resource state fluctuations, and complex task dependencies in dynamic heterogeneous distributed computing environments, this paper proposes an intelligent task scheduling method oriented towards structure awareness and long-term decision optimization. This method models the distributed scheduling process as a unified sequential decision problem. By constructing a heterogeneous graph representation involving both task nodes and resource nodes, it incorporates pre-task and post-task constraints, node differences, and interactions between tasks and resources into a single modeling framework. Building upon this, a graph neural network is used to deeply represent the system state, enhancing the model's understanding of complex topological relationships and dynamic runtime contexts. Furthermore, deep reinforcement learning is integrated to learn scheduling strategies, enabling the model to formulate more rational task allocation schemes from a global benefit perspective. Simultaneously, to improve the adaptation quality between tasks and resources, a task-resource matching module is designed to characterize the correlation strength between executable tasks and candidate computing nodes, and a joint objective optimization approach coordinates the structure representation learning and strategy optimization processes. Related research shows that the proposed method effectively improves the state representation capability, decision consistency, and resource allocation rationality in the scheduling process, providing a unified solution with structural modeling and adaptive decision-making capabilities for intelligent scheduling in complex distributed scenarios.