Title: Optimizing Dynamic Cloud Workflow Scheduling with Variable Deadlines Using a Mixture-of-Experts Approach

Abstract:

In the realm of cloud computing, workflow scheduling requires the intelligent distribution of graph-structured tasks, which arrive dynamically and carry distinct deadline constraints, across fluctuating virtual machine resources. Current deep reinforcement learning (DRL) schedulers, however, are constrained by inflexible, single-path inference structures that often fail to cope with the variety of scheduling contexts. To address this, we present DEFT (Deadline-pErceptive Mixture-oF-Experts), a novel DRL policy framework that employs a specialized mixture of experts, where each expert is calibrated to handle specific degrees of deadline pressure. As far as we know, DEFT marks the first introduction and validation of a Mixture-of-Experts architecture within the domain of dynamic cloud workflow scheduling. By dynamically channeling decisions to the most suitable experts, DEFT successfully satisfies a wide array of deadline demands that would be unattainable for any individual expert operating in isolation. The core of DEFT is a graph-adaptive gating mechanism that utilizes cross-attention to encode workflow deadlines, directed acyclic graphs (DAGs), task statuses, and virtual machine conditions. This mechanism guides expert activation with fine-grained sensitivity to deadlines. Benchmark tests on dynamic cloud workflows reveal that DEFT substantially lowers execution costs and minimizes deadline violations, surpassing various state-of-the-art DRL baseline methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC