Research On The Method Of Multi-VM Resource Adaptive Adjustment In A Single PM For The Concurrent User Request

With the development of cloud computing, virtual machines (VMs) are widely used in real life, and the problem of wasting resources and interfering service performance faced by large servers have been solved preliminarily. However, the traditional method of VMs resource allocation is static allocation, that is to say the initial amount of resources allocation of the VM is "lifelong". The amount of resources will not be changed during the time VMs are running. This method of resource allocation may lead to two results:(1) Excessive resources allocation, VMs will never use this amount of resources allocation; (2) Inadequate resources allocation, the phenomenon of VMs' resources shortage will often appear. How to allocate resources to VMs becomes a serious problem.To solve the problems caused by the method of static resource allocation, the thought of dynamically allocate VMs resources came into our life. The life cycle of VMs can be divided into several resource adjustment cycle. Allocate a fixed amount of resources in each VM resource adjustment cycle, that is to say "static in cycle, and dynamic out cycle". This can ease the phenomenon of wasting resources on the premise of guaranteeing the quality of service. In this thesis, facing to the problem that how to dynamically allocate multi-VM resource in a single physical machine(PM) on the premise of service performance according to the number of concurrent user request, propose the method that multi-VM resources adaptive adjustment in a single PM for the concurrent user request. Firstly, analysis the historical data produced during the time VMs are running, and build the relation model between the number of concurrent user request and the amount of resources consumption (C-R Model) and the relation model between the service performance and the amount of resources consumption (P-R Model). Secondly, use the BP neural network algorithm to predict the number of concurrent user request and compute the average of the number of concurrent user request in the next adjustment cycle. Then use the average and the C-R Model to assess the amount of resource demand by VMs in the next adjustment cycle. And then use genetic algorithms (GA) and the P-R Model to evaluate the performance of the individual each generation. Then choose the best individual, and put the individual selected as well as the evaluation of the implementation into the experience database. Finally, use the experiment to proof the effectiveness, the feasibility and the superiority of the method proposed in this thesis.Experimental results show that the method proposed in this thesis can effectively evaluate the amount of resources demand by VMs. And at the time determine the initial population of GA, its range can cover all possible solutions. Build experience database also can save a lot of computing time when choose the optimal solution. So it become the key to complete optimal decision task before the deadline.