Research On Collaborative Computation Offloading In Multi-Access Edge Computing

The rapid development of Internet of Things(IoT)technology gives birth to a large number of advanced applications with delay-sensitive and computation-intensive characteristics.In order to achieve lower energy consumption,faster response,greater security,higher reliability and less bandwidth requirements,Multi-access Edge Computing(MEC)technology emerges and gains its popularity.However,lack of unified scheduling management among edge server in traditional edge computing paradigm leads to computing power island effect,while ubiquitous computing resource cannot be fully leveraged.Therefore,how to build a unified computing power management and awareness system,design a collaborative computation offloading strategy for diverse requests,achieve extremely flexible computation offloading and extremely flexible computing resource allocation has become a hot spot in academic research.Based on this,this topic starts from several typical application scenarios of multi-access edge computing,proposes collaborative computation offloading strategies from the perspective of centralized and distributed management.Besides,we designed and implemented a collaborative computing offloading system based on microservices architecture based on the proposed offloading strategy,in order to verify the comprehensive performance of the scheduling strategy in real edge inferring environments.Specifically,the main contributions of this paper include the following three aspects:Firstly,we consider the scenario of multi-user devices with multiple edge computing servers,and study the joint optimization of computation offloading and resource allocation with the goal of minimizing the energy consumption.We design apriority-based resource allocation strategy when the resource is limited,transform the computation offloading problem into a weighted bipartite graph matching and minimum-cost maximum-flow problem,and propose KM(Kuhn-Munkres)algorithm as well as improved GS(Gale-Sherply)algorithm to solve the matching problem.The simulation results show that the proposed MADF-GS(Maximum Alternative Differences First GS)algorithm has excellent performance in comparison to the traditional random scheduling(RAO)strategy and the theoretically optimal KM scheduling strategy,whose performance is stable when hyperparameter changes,and the network energy consumption is nearly optimal when the node density increases,61.21%higher than RAO strategy.Compared with the KM strategy,the running time of is MADFGS can save up to 80%.Secondly,considering the highly dynamic network environment,we select the Internet of Vehicles as the research scenario,comprehensively measure the backlog of task queues and the amount of offloading data,and study the dynamic collaborative computation offloading problem between vehicles and vehicles,vehicles and road side units(RSUs).In order to solve the problem of data transmission when vehicles offload computing tasks to non-direct-connected drive RSUs,we design a fast greedy path finding strategy suitable for distributed scenarios;In addition,we transform the problem of dynamic collaborative computation offloading as a game of dynamic resource pricing.By considering the game between offloading data and dynamic resource pricing,we obtain the Stackelberg equilibrium solution based on careful mathematical derivation.The simulation results reveal that compared with the classic A*algorithm,the fast greedy routing strategy greatly reduces the execution time of the algorithm while the performance gap is no more than 1%.Compared to various traditional priority-based greedy offloading strategies,the Stackelberg game based dynamic collaborative computation offloading strategy can balance the offloading data and resource pricing while ensuring the stability of the task queue.Compared with the most high-performance greedy algorithm,it can achieve a 40.94%improvement in system utility.Thirdly,we have designed and implemented a complete collaborative computation offloading system based on the microservice architecture.We model the request scheduling problem in actual environment,and design a scheduling strategy suitable for the actual network environment based on the collaborative computation offloading strategy proposed in research point 1;We introduce the container cluster management platform to uniformly manage edge computing nodes,and design several microservices such as computing power awareness,computing power announcement,computing power modeling and request scheduling to realize flexible and dynamic scheduling of requests.Our tests show that in diversed edge inferring services scenario,the computing collaborative offloading system ensures balanced resource utilization while reducing response latency of 69.51%,17.91%and 20.14%compared with typical traditional paradigm including local,random,and greedy strategy,and achieves rapid response as well as optimal throughput under the dynamic network topology,which verifies the technical feasibility and performance superiority of collaborative computation offloading.