HiGFRL: Hierarchical Graph Fusion-Driven Reinforcement Learning for Dependency-Aware Task Scheduling in Heterogeneous Cloud

Summary

The paper proposes HiGFRL, a deep reinforcement learning framework for online scheduling of dependency-aware (DAG) tasks on heterogeneous virtual machine clusters. It targets two claimed deficiencies of prior GNN-based schedulers: weak modeling of many-to-one join dependencies and loose coupling between topology and resource state. HiGFRL builds a three-view state — a static task-dependency hypergraph whose hyperedges group each task with all predecessors, a dynamic global cluster graph with masked pooling, and a local task-VM bipartite matching graph — processed by a PPO-style dual-stream actor-critic (Context Fusion Allocator plus Global State Evaluator) with hard action masking and a hybrid reward adding an earliest-finish-time term and an upward-rank topological prior. On Alibaba 2018 cluster traces across six workload scenarios (300-900 jobs, 40 VMs), the authors report the best makespan, average flow time, and wait time versus Random, Round-Robin, Greedy-Tetris, READYS, GA-DRL and MODRL, plus mostly the lowest retry counts; headline gains are a 32.55% makespan reduction (Large-100: 1,198,246 vs 1,776,491), and 13.58% flow-time and 13.79% wait-time improvements (Large-250). An ablation removing the hierarchical fusion degrades to a flat GAT and loses most large-scale gains (Table 5). Limitations: all results are single-run point comparisons without variance or significance tests, gains may partly stem from the heuristic-shaped reward, baseline tuning and hyperparameters are thinly documented, and evaluation is simulation-only.