Research On Performance Optimization In Edge Computing With Stochastic Factors

The rapid development of mobile Internet and Internet of things(IoT)technology has hastened many computational intensive and time-delay-sensitive applications,such as cloud game,auto-driving,live video,face recognition,etc.These applications not only increase the computing burden of User Equipment(UE)and IoT devices,but also generate a large volume of data transmission to backbone network and cloud data center,and these applications are not suitable to be deployed in cloud computing center due to the large time delay caused by the wide area network(WAN)transmission.Edge computing is a distributed service computing mode that deploys computing and storage resources to the user's near end.It can provide low delay computing and storage services for UE or IoT nodes,help to reduce the data transmission in backbone network,and reduce the pressure of data transmission and processing in cloud centers.In addition,the UE or IoT device can migrate its computing services to the edge server side,which increases the computing capacity at the same time reduces its own computing power consumption.Because of the advantages mentioned above,edge computing has become a hot topic in the field of service computing,which has attracted extensive attention from both academia and industry.In edge computing,the related performance indicators includes delay,energy,and reliability,and optimizing the above indicators have significant importance in both theory and practice.Existing work on workload scheduling and resource allocation ignored the impact of random fluctuation of workload and resource demand,and assumes that the workloads are fixed valued.In addition,when considering the delay cost,it is generally considered that the delay is fixed,or only consider the average delay cost,and delay jitter is not considered.By considering the characteristics of time delay and random fluctuation of service resource demand,this thesis studied the scheduling and resource allocation of edge computing service from the perspective of service reliability,and the main contributions are summarized as follows:(1)A delay jitter-resistant multi-access edge computing task scheduling algorithm is proposed.In the multi-access edge computing scenario,task offloading scheduling problem considering task call graph is a typical task scheduling problem in parallel het-erogeneous computing.In this thesis,two new delay time indator,namely Delay Risk Probability(DRP)and Maximum Tolerable Delay(MTD)are proposed,and the cor-responding task scheduling model is constructed to minimize MTD.On the basis of well-known Heterogeneous Earliest-finish-time(HEFT)algorithm,an improved Conser-vative CHEFT(CHEFT)and its Gaussian approximated Gau-CHEFT Algorithm are proposed,in which both average delay time and delay-jitter are considered.Simulation results show that the proposed CHEFT and Gau-CHEFT algorithms can achieve lower MTD compared with HEFT algorithm and greedy algorithm,which demonstrates that it has stronger anti-delay jitter capability.(2)A workload balance and dynamic resource configuration based energy minimiza-tion algorithm is proposed for cooperative edge computing with stochastic workloads.By jointly consider the computation energy cost and reconfiguration energy cost,a new energy cost model is proposed modeling the energy cost of VM servers.Theoretically,it is proved that the expected energy consumption has a unique minimum point,and the condition of obtaining the minimum energy consumption value is derived,on the basis of this condition,an optimal resource configuration algorithm based on bisection search method is proposed.Furthermore,the workload balancing and dynamic capac-ity configuration in multi-access edge computing are studied.When the distribution of the workloads is not specified,its PDF is estimated by using Histogram,and a heuristic search method is proposed to find the workload assignment vector(WAV).When the workloads follow Gaussian distribution,an alternative optimization method is proposed to solve the problem by decomposing the primal problem into a workload assignment subproblem and a resource configuration subproblem,the former one is a convex problem that can be efficiently solved by classical convex method,the latter one is a quasiconcave problem,and can be optimally solved by the proposed bisection method.The results of both Monte Carlo Simulation and Planetlab data verify the validity of the proposed energy consumption model,and show that the proposed dynamic configuration algorithm can achieve lower energy cost compared with the static resource configuration method.(3)A service reliability maximization algorithm based on Quality of service(QoS)control and resource allocation is proposed.This thesis defines Service Reliability Prob-ability(SRP)as the probability that the real-time computing resource requirement of a server or VM is not larger than its processing capacity.Furthermore,the Average SRP(ASRP)maximization problem for VM architecture and the System SRP(SSRP)maximization problem for container architecture are developed.In order to solve the ASRP maximization problem,an alternative optimization algorithm is proposed,which decomposes the original problem into Resource Allocation Problem(RAP)and Service Quality Control Problem(SQCP),the former is firstly transformed to a convex problem by using Logistic approximation,and then it is solved by using KKT conditions and the primal-dual method.The latter is a quasi-convex problem,and it is solved by the bi-section method and second-order cone programming.The SSRP maximization problem under container architecture is naturally the same with SQCP problem,which can be solved in the same way.The simulation results show that the proposed algorithm can improve the service reliability,and demonstrate that the container architecture is better than VM architecture in improving service reliability.(4)A service reliability maximization algorithm based on multi-dimensional resource allocation is proposed.The resources involved in edge computing include computing,in-put/output(I/O),memory,disk read/write,bandwidth,to name a few.Based on the modeling of service reliability for VM architecture and container architecture,the service reliability maximizing problem based on multi-dimensional resource allocation are devel-oped.For the one in VM environment,an alternate optimization method is proposed,which optimizes the resource allocation of one dimension in each alternate iteration,and then uses Logistic approximation to transform the objective function into a concave function,which can be efficiently solved by classical convex methods.For the problem in container environment,it can be transformed into a quasi-convex problem only with range constraints.By using logarithmic transformation,the objective is transformed into a quasi-concave function,which can be solved directly by solving the first derivative.The simulation results show that the proposed method can achieve better service reliability than the baseline methods under both VM and container server architecture,and gain demonstrate that container architecture is better than VM architecture in improving service reliability.