文献类型：期刊文献

英文题名：Adaptive edge service deployment in burst load scenarios using deep reinforcement learning

作者：Xu, Jin[1];Yu, Huiqun[1,2];Fan, Guisheng[1];Zhang, Jiayin[1];Li, Zengpeng[1];Tang, Qifeng[3]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China;[2]Shanghai Key Lab Comp Software Evaluating & Testin, Shanghai 201112, Peoples R China;[3]Shanghai Data Exchange Corp, Shanghai 200436, Peoples R China

年份：2024

卷号：80

期号：4

起止页码：5446

外文期刊名：JOURNAL OF SUPERCOMPUTING

收录：;EI(收录号：20234014828370);WOS:【SCI-EXPANDED(收录号:WOS:001074833500002)】；

基金：This work was partially supported by the Natural Science Foundation of Shanghai (No. 21ZR1416300), the Capacity Building Project of Local Universities Science and Technology Commission of Shanghai Municipality (No. 22010504100), the Research Programme of National Engineering Laboratory for Big Data Distribution and Exchange Technologies, and the Shanghai Municipal Special Fund for Promoting High Quality Development (No. 2021-GYHLW-01007).

语种：英文

外文关键词：Burst load; Edge service; Scaling and migrating; Deep reinforcement learning

摘要：The development of edge computing provides a novel deployment strategy for delay-aware applications, in which applications initially deployed in central servers are shifted closer to end-users for higher-quality and lower-delay services. However, with the growth in the number of end-users and devices, edge services are increasingly susceptible to sudden load spikes. In burst load scenarios, deploying services and allocating resources to maintain service quality and load balancing of edge servers become challenging, particularly given the coupling of resource requirements between services. This paper addresses this challenge by modeling the load burst scenario as a Markov decision problem and proposing a deep reinforcement learning-based (DRL-based) approach. The proposed approach ranks services based on their migration status and request delay violations, and makes scaling and migration decisions for each service in turn, with the goal of maximizing the total request throughput while satisfying delay requirements and resource constraints. Simulation results show that the proposed approach outperforms other algorithms in terms of total throughput and delay violation rate.