面向移动边缘的组合服务选择及优化

移动边缘计算作为新型的计算范式,为降低网络延迟、能耗开销提供了新的思路.其将中心云的强大算力下沉至网络边缘,使得用户能够将计算任务卸载至物理位置更近的边缘服务器执行,从而节省经由核心网的时延与能耗开销.然而,由于移动边缘计算技术通常受到计算资源、网络传输带宽、设备电量等因素的制约,如何在有限的资源中获取最大的利用率成为亟待解决的难题.此外,复杂的网络服务可以被抽象为由若干个子服务按照一定拓扑结构组成的组合服务,然而紊乱多变的移动网络环境为用户策略赋予了时空特性、决策耦合、边缘节点异构以及计算复杂度高的特性,使得传统的基于QoS(Quality of Service)的算法不再适用.本文建立由异构边缘节点以及装配有能量收集组件的移动设备组成的移动边缘系统,基于李雅普诺夫优化以及马尔科夫近似提出一种多项式计算复杂度的分布式算法,提出CSS(Composite Service Selection)框架,旨在联合优化服务选择策略以及能量存储策略,以此最小化整体组合服务请求的总体响应时间,并将设备电量稳定在一个可靠的水平.本文选取四种基准算法,实验结果表明CSS框架具备更加良好的性能,在时延上优于其他算法7.76%~28.88%,并能够最快实现电量稳定.随着场景规模的扩大,CSS将体现更优的性能.

Composite Service Selection and Optimization for Mobile Edge Systems

Mobile edge computing has emerged as the promising paradigm, devoted to provide innovative ideas in reducing latency and saving energy consumption. It enables users to offload computation tasks to edge servers in the proximity by pushing powerful functionalities of central cloud to network edge, so as to reduce overhead spent on transmission via core network, in terms of latency and energy. However, due to real-life restrictions, such as computing resources, caching capabilities and network transmission bandwidth of edge servers, edge nodes could only provision stringent-limited services for users. Moreover, battery capability of user equipment is hardware-constrained. It is crucial to explore the potential of obtaining full utility under insufficient resources with mobile edge computing technology. With the concept of microservice proposed, polybasic network services can be abstracted as composite services consist of multiple subservices with complex structure. As mobile devices evolved to be capable of processing more diversified mobile services, people tend to exploit it to handle daily business. On the other hand, network state in mobile environment is volatile and location-based which endows spatial and temporal properties on users’ service-oriented strategies, making traditional QoS-aware algorithm is not applicable any more. Addressing this problem faces challenges of decision coupling, edge node heterogeneity and high computation complexity. Most existing work does not take into consideration all of the above factors. To bridge this gap, in this paper, we consider a mobile edge system composed of heterogenous edge nodes and mobile users who generate composite services requests randomly at the begin of each time slot and are equipped with energy harvest components for sustainable computing with external energy conversion. To optimize latency and guarantee that battery power is maintained at a stable and reliable level, we have to determine the selection and offloading strategies of composite service requests and energy harvesting policies for mobile devices in each time slot, which has non-polynomial complexity. We formulate it as a stochastic optimization problem and proposes a new framework named CSS(Composite Service Selection) aiming at minimizing the overall response time of composite service requests generated by users in a mobile community and stabilizing battery energy in a reliable level, with the above constraints considered. First, it transforms the stochastic optimization problem over time slots into normal optimization problem in a given time slot, by introducing Lyapunov optimization. Then this problem could be decoupled into two subproblems as energy harvest problem and service selection problem. The closed-form expression of the energy harvest problem can be derived directly and next we develop a distributed method to solve the service selection problem, based on Markov approximation which has polynomial computation complexity. We take four baseline algorithms as benchmarks which are adapted from previous works. The experimental results demonstrate the superiority of our proposed framework compared with other baseline algorithms based on real-world dataset. It can obtain asymptotic optimality with low complexity and outperform other algorithms from 7.76% to 28.88% in term of latency. Moreover, mobile devices utilizing CSS are the fastest to reach stability. As problem scales, CSS outperforms other algorithms from 3.7% to 55.24%.