Deft Scheduling of Dynamic Cloud Workflows with Varying Deadlines via Mixture-of-Experts

Researchers introduce DEFT, a new deep reinforcement learning architecture using a mixture-of-experts approach to optimize cloud workflow scheduling with varying deadline constraints. The system uses a graph-adaptive gating mechanism to route scheduling decisions through specialized experts, demonstrating improved performance in reducing execution costs and deadline violations compared to existing DRL baselines.

DEFT represents a meaningful advancement in cloud computing infrastructure optimization by addressing a fundamental limitation of current deep reinforcement learning schedulers. Traditional DRL systems rely on single-path inference architectures that struggle when handling diverse scheduling scenarios with different deadline pressures, leading to suboptimal resource allocation decisions. The introduction of a mixture-of-experts framework allows the system to specialize different neural network components for specific deadline contexts, enabling more nuanced decision-making.

The development builds on established trends in both machine learning and cloud infrastructure management. As cloud computing workloads become increasingly complex and time-sensitive, schedulers must balance competing objectives like cost minimization and deadline adherence. Previous DRL approaches attempted to handle this diversity within a single model, inherently limiting their effectiveness. The mixture-of-experts paradigm has proven successful in other domains, and its application to workflow scheduling represents a logical progression.

For cloud infrastructure operators and enterprises, DEFT offers practical benefits through reduced operational costs and improved service reliability. The graph-adaptive gating mechanism, which encodes workflow structure and machine conditions, enables fine-grained adaptation to real-time constraints. This specificity contrasts with rigid scheduling policies that often over-provision resources to guarantee deadline compliance, wasting capital.

The research opens opportunities for further optimization in distributed computing. Future work might explore hierarchical mixture-of-experts structures for extremely complex workloads or integration with heterogeneous computing resources including GPUs and specialized hardware. The validation on dynamic benchmarks suggests production readiness, potentially influencing how major cloud providers architect their scheduling layers in coming years.

• →DEFT introduces the first mixture-of-experts architecture specifically designed for dynamic cloud workflow scheduling with deadline constraints.
• →The graph-adaptive gating mechanism uses cross-attention to route decisions based on workflow structure and deadline tightness, enabling specialized expert handling.
• →Experimental results demonstrate significant reductions in execution costs and deadline violations compared to state-of-the-art DRL baselines.
• →The approach overcomes rigid single-path inference limitations by allowing different experts to specialize in handling varying deadline scenarios.
• →The system dynamically encodes DAG structure, task states, and VM conditions to make context-aware scheduling decisions.