Cloudlet-based Mobile Cloud Platform Workload Distribution And Resource Allocation

With the development of Internet technology,the applications of various functions are constantly being produced,and the requirements for delay and bandwidth have become stricter.Using MCC technology to migrate computationally intensive parts to cloud is a common way to improve user experience.However,data center has high latency and low bandwidth,which is not suitable for delay-sensitive interactive applications.The proposal of cloudlet solves this problem.Cloudlet is a small cloud data center distributed at the edge of the network,providing users with low latency and high bandwidth services.Different application scenarios are different,so the importance and latency requirements of the application are different.How to reduce the resource usage of cloud infrastructure providers(CIP)and ensure the execution of key applications without reducing the application's QOS is a valuable problem.From the perspective of CIP,this dissertation aims to optimize the SLA violations aware cost,and research the workload distribution and resource allocation problem for task and service.For service application,the cost of the CIP is related to request distribution and service deployment.Service deployment number determines the resource cost of the CIP.Service deployment location and request distribution determine whether the user request violates the time threshold.Suitable request distribution and service deployment solutions can effectively reduce cost.This dissertation proposes a cost-optimized greedy algorithm named CO-Greedy,which can find a suitable request distribution and service deployment solution in polynomial time.CO-Greedy algorithm will prioritize the deployment of more important services,and when cloudlet resource usage exceeds the preset threshold,it can migrate the service container with the lowest gain,reducing the cost of CIP.This dissertation tests the performance of CO-Greedy and the existing three algorithms under different time thresholds,different resource scale and different request scale through simulation experiments,which proves that the CO-Greedy algo-rithm provides lower cost.For task application,the cost of the CIP is related to task distribution and resource allocation.Task resource allocation not only determines the resource cost of the CIP,but also determines the execution speed of the task,thereby affecting whether the task violates the time threshold.Finding the suitable position for the task can improve cloudlet utilization and reduce the probability of the task violating the time threshold.In addition,multiple users continuously access the cloudlet platform,and each user's decision will affect the later users,so task distribution and resource allocation problem is a sequence decision problem.This dissertation proposes a reinforcement learning based task distribution and resource allocation algorithm named RLTDRA.This algorithm performs Markov Decision Process(MDP)modeling on task distribution and resource allocation problems,and builds a value network using Deep Q Networc(DQN),and continuously updates parameters to find low-cost task distribution and resource alloca-tion schemes.This dissertation tests the performance of RLTDRA algorithm and online greedy algorithm under different user access rate and resource scale through simulation experiments,and proves that the cost of scheme provided by RLTDRA algorithm is lower.