A Learning Method with Gap-Aware Generation for Heterogeneous DAG Scheduling
Researchers have developed WeCAN, a novel reinforcement learning framework designed to optimize the scheduling of directed acyclic graphs (DAGs) in large-scale computing systems. This framework addresses challenges such as task-pool compatibility and optimality gaps introduced by generation processes. WeCAN employs a two-stage single-pass design that generates task-pool scores and global parameters, followed by a schedule construction map. Experiments on TPC-H query DAGs and ML-compiler computation graphs show that WeCAN outperforms existing baselines in makespan while maintaining competitive inference times. AI