A Learning Method with Gap-Aware Generation for Heterogeneous DAG Scheduling

arXiv cs.LG / 3/25/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

Key Points

  • The paper introduces WeCAN, a reinforcement-learning framework for efficient scheduling of heterogeneous DAGs that accounts for task–pool compatibility coefficients and mitigates “generation-induced optimality gaps.”
  • WeCAN uses an end-to-end, two-stage single-pass design: one forward pass computes task–pool scores and global parameters, then a generation map constructs schedules without repeated network calls.
  • The approach employs a weighted cross-attention encoder to model task–pool interactions in a way that is size-agnostic to changes in the environment’s resource pools and task types.
  • It presents an order-space analysis to characterize which schedule orders are reachable by the generation map, explaining how restricted reachability creates suboptimality gaps.
  • By using sufficient conditions derived from the analysis, the authors design a skip-extended generation rule that enlarges the reachable order set while keeping inference comparable to classical heuristics and faster than multi-round neural schedulers.

Abstract

Efficient scheduling of directed acyclic graphs (DAGs) in heterogeneous environments is challenging due to resource capacities and dependencies. In practice, the need for adaptability across environments with varying resource pools and task types, alongside rapid schedule generation, complicates these challenges. We propose WeCAN, an end-to-end reinforcement learning framework for heterogeneous DAG scheduling that addresses task--pool compatibility coefficients and generation-induced optimality gaps. It adopts a two-stage single-pass design: a single forward pass produces task--pool scores and global parameters, followed by a generation map that constructs schedules without repeated network calls. Its weighted cross-attention encoder models task--pool interactions gated by compatibility coefficients, and is size-agnostic to environment fluctuations. Moreover, widely used list-scheduling maps can incur generation-induced optimality gaps from restricted reachability. We introduce an order-space analysis that characterizes the reachable set of generation maps via feasible schedule orders, explains the mechanism behind generation-induced gaps, and yields sufficient conditions for gap elimination. Guided by these conditions, we design a skip-extended realization with an analytically parameterized decreasing skip rule, which enlarges the reachable order set while preserving single-pass efficiency. Experiments on computation graphs and real-world TPC-H DAGs demonstrate improved makespan over strong baselines, with inference time comparable to classical heuristics and faster than multi-round neural schedulers.