一种基于深度强化学习与概率性能感知的边缘计算环境多工作流卸载方法

Multi-workflow Offloading Method Based on Deep Reinforcement Learning and Probabilistic Performance-aware in Edge Computing Environment

移动边缘计算是一种新兴的分布式和泛在计算模式,其将计算密集型和时延敏感型任务转移到附近的边缘服务器,有效缓解了移动终端资源不足的问题,显著减小了用户与计算处理节点之间的通信传输开销。然而,如果多个用户同时提出计算密集型任务请求,特别是流程化的工作流任务请求,边缘计算环境往往难以有效地进行响应,并会造成任务拥塞。另外,受任务负载、电力供给、通信能力的实时变化等不利因素的影响,边缘服务器本身的性能总是处于波动和变化中,从而为保证任务执行效能和用户感知服务效率带来了挑战。针对上述问题,文中提出了一种基于深度Q网络(DQN)与概率性能感知机制的边缘计算环境多工作流调度方法。首先对边缘云服务器的历史性能数据进行概率分析,然后利用获得的性能概率分布数据驱动DQN模型,不断迭代优化,生成多工作流的卸载策略。在实验验证环节,基于边缘服务器位置数据集、性能测试数据和多个科学工作流模板,在反映不同系统负载水平的多个场景下进行了模拟实验。实验结果表明,所提方法在多工作流执行效率方面明显优于传统方法。

Mobile edge computing is a new distributed and ubiquitous computing model.By transferring computation-intensive and time-delay sensitive tasks to closer to the edge servers,it effectively alleviates the resource shortage of mobile terminals and the communication transmission overhead between users and computing processing nodes.However,if multiple users request computation-intensive tasks simultaneously,especially process-based workflow task requests,edge computing are often difficult to respond effectively and cause task congestion.Inaddition,the performance of edge servers is affected by detrimental factors such as task overload,power supply and real-time change of communication capability,and its performance fluctuates and changes,which brings challenges to ensure task execution and user-perceived service efficiency.To solve the above problems,a Deep-Q-Network(DQN)and probabilistic performance aware based multi-workflow scheduling approach in edge computing environment is proposed.Firstly,the historical performance data of edge cloud servers is analyzed probabilistically,then the DQN model is driven by performance probability distribution data,and iterative optimization is carried out continuously to generate multi-workflow offloading strategy.In the process of experimental verification,simulation experiments are conducted in multiple scenarios reflecting difterent levels of system load based on edge server Location data set,performance test data and multiple scientific workflow templates.The results show that the proposed method is superior to the traditional method in the execution efficiency of multi-workflow.