Abstract:

In Computing Power Network （CPN）， User Device （UE） with limited computational capability and battery capacity rely on external computing nodes for collaborative task processing. The existing study mainly focuses on direct Workflow Task （WT） offloading， but the following key challenges are faced： 1） long waiting latency and high energy consumption caused by task dependencies； 2） high-power mode maintained by UE persistently when precursor task data required to be cached on UE； 3） complexity of offloading decisions increased by uncertainties in resource states within CPN dynamic environments； 4） the difficulty in achieving efficient balance caused by multi-objective conflicts between task completion time and energy consumption. To address these challenges， a Dynamic Optimization and Offloading for Workflow Task （DOOWT） was developed to improve energy efficiency. In the algorithm， the Workflow Structure Optimization （WSO） algorithm was utilized to rearrange the task graph， so as to reduce task waiting latencies， thereby lowering overall energy consumption； a Dynamic-Based Task Offloading （DBTO） algorithm based on Deep Deterministic Policy Gradient （DDPG） was employed to adjust offloading strategies dynamically， thereby enhancing computational performance and resource utilization in CPN. Experimental results demonstrate that compared with conventional methods such as Random unloading （Random）， the proposed method reduces the WT waiting latency by 60%， shortens the average WT completion latency by 79%， and decreases the overall energy consumption by 82%. It can be seen that this method provides theoretical and technical support for the optimization and scheduling of energy consumption-sensitive tasks.

Key words: Computing Power Network (CPN), Dynamic-Based Task Offloading (DBTO), Deep Reinforcement Learning (DRL), energy consumption optimization, task offloading

摘要：

在算力网络（CPN）中，用户设备（UE）因计算能力和电源容量受限，需要依赖外部算力节点协同处理任务。现有研究多集中于直接卸载工作流任务（WT）上，而面临着以下关键挑战：1）任务依赖导致的长等待时延和高能耗问题；2）当前驱任务数据需在UE上缓存时，UE长时间处于高功耗状态；3）CPN动态环境下资源状态的不确定性增加了卸载决策的复杂性；4）任务完成时间与能耗之间的多目标冲突，导致难以实现高效平衡。针对这些问题，提出一种基于动态任务优化与卸载（DOOWT）的节能优化方法。该算法通过工作流结构优化（WSO）算法对任务图进行重排，以减少任务间的等待时延，从而降低整体能耗；结合基于深度确定性策略梯度（DDPG）的动态任务卸载（DBTO）算法实时调整卸载策略，从而有效提升CPN的计算性能和资源利用效率。实验结果表明，与随机卸载（Random）传统方法相比，所提方法的WT等待时延减少了60%，平均WT完成时延缩短了79%，整体能耗降低了82%。可见，该方法为能耗敏感型任务的优化与调度提供了理论支持与技术参考。

关键词: 算力网络, 动态任务卸载, 深度强化学习, 能耗优化, 任务卸载