PF-MPPO: Task-dependent workflow scheduling method based on deep reinforcement learning in dynamic heterogeneous cloud environments

Donghong Xu

doi:10.1016/j.future.2025.108236

PF-MPPO: Task-dependent workflow scheduling method based on deep reinforcement learning in dynamic heterogeneous cloud environments

2025-11-08

0

PRE

AI

DOI：10.1016/j.future.2025.108236

Original

Original request for help

Save

Abstract

En 中文

Workflow task scheduling in heterogeneous dynamic cloud environments faces significant challenges, including balancing task precedence constraints with parallel execution and adapting to dynamic node configurations. This paper proposes PF-MPPO (Pre-training Fine-tuning Multi-agent Proximal Policy Optimization), a Deep Reinforcement Learning (DRL)-based algorithm designed for such environments. PF-MPPO integrates pre-training and fine-tuning within a multi-agent architecture, modeling workflows as Directed Acyclic Graphs (DAGs) and optimizing for latency, energy consumption, and load balancing. A dual-stage PPO framework enables rapid adaptation by comparing current task clusters with pre-trained models and selecting the most suitable one for incremental learning. Simulation results demonstrate that PF-MPPO outperforms baseline algorithms, significantly reducing latency and energy consumption while maintaining load balancing, validating its effectiveness in dynamic cloud environments.

PF-MPPO: Task-dependent workflow scheduling method based on deep reinforcement learning in dynamic heterogeneous cloud environments

Abstract

Journal

Organization