Abstract:

Edge Computing (EC) provides users with low-latency, high-response services at the edge of a network. Therefore, task offloading strategies with high-resource utilization and low-latency have become a popular research direction. However, most existing task offloading research is based on a centralized architecture, and offloading strategies and resource scheduling are conducted using centralized facilities, which produce additional energy consumption, high-latency, and are susceptible to the risk of a single point of failure. To address these challenges, a Decentralized Priority Deep Q Network (DP-DQN) offloading strategy is proposed. First, a communication matrix is established to simulate the limited communication state of the edge server. Second, by setting task priorities, tasks can be switched between different edge servers to ensure that each server independently formulates uninstallation policies, to complete decentralization of the uninstallation task. Finally, additional computing resources are allocated to tasks based on the number of jumps to improve resource utilization efficiency and the optimization effect. To verify the effectiveness of the proposed strategy, a comparative study is conducted on the convergence performance of parameters under different DQN. The experimental results demonstrate that in different testing scenarios, the performance of the proposed DP-DQN strategy is superior to that of local, fully greedy, and multi-objective task offloading algorithms, and the performance could be improved by approximately 11%-19%.

Key words: Edge Computing(EC), task offloading, resource allocation, decentration, priority, Deep Q Network(DQN)

摘要：

边缘计算(EC)可在网络边缘为用户提供低延迟、高响应的服务。因此, 资源利用率高、时延低的任务卸载策略成为研究的热门方向。但大部分现有的任务卸载研究是基于中心化的架构, 通过中心化设施制定卸载策略并进行资源调度, 容易受到单点故障的影响, 且会产生较多的能耗和较高的时延。针对以上问题, 提出一种基于深度Q网络(DQN)的去中心化优先级(DP-DQN) 卸载策略。首先, 设置通信矩阵模拟现实中边缘服务器有限的通信状态; 其次, 通过对任务设定优先级, 使任务可以在不同边缘服务器之间跳转, 保证各边缘服务器均可以自主制定卸载策略, 完成任务卸载的去中心化; 最后, 根据任务的跳转次数为任务分配更多的计算资源, 提高资源利用效率和优化效果。为了验证所提策略的有效性, 针对不同DQN下参数的收敛性能进行了研究对比, 实验结果表明, 在不同测试情景下, DP-DQN的性能均优于本地算法、完全贪婪算法和多目标任务卸载算法, 性能可提升约11%~19%。

关键词: 边缘计算, 任务卸载, 资源分配, 去中心化, 优先级, 深度Q网络