Resource Deployment And Optimization In Datacenter Based On Queuing Theory And BP Neural Network

With the rapid development of software defined network(SDN)and network function virtualization(NFV),the existing problems in traditional data center,such as low resource utilization,difficulty of extending network functions and high operating costs,gradually find solutions.This paper,which under the background of NFV,designs a dynamic resource deployment and optimization strategy based on the queuing theory and back-propagation neural network,aiming at the low resource utilization problem.The main innovation points of this paper are as follows:(1)We design a simulated generator of service function chain(SFC),which considers demand for service.We set the arrival rate of SFC and the service rate of service function according to the queuing theory,beyond that,we combine the demand ratio for all kinds of service function to generate the SFC flow in datacenter.(2)We design the parallel architecture of service system and dynamic resource deployment strategy.In the designed service system,each server integrates all the service functions,which can not only improve the system fault tolerance,but also avoid the bandwidth resource consumption among servers.At the same time,we dynamically schedule resource of the virtual machines in one sever on demand,maximizing the utilization of the shared resource.(3)We use the back-propagation(BP)neural network based on the L-BFGS algorithm to predict the resource utilization,and then optimize the resource configuration according to the predicted values.We regard the arrival rate of SFC,service rate of service function and system resource configuration as features of data samples,and resource utilization as the label of data samples.Then train and test the BP neural network based on L-BFGS with simulated data samples.Upon the predicted resource utilization,we can find out the bottleneck resource and redundant resource,which we can optimize respectively and further increase the utilization.In this article,we make comparing experiments with our dynamic resource deployment,which under service function integration and parallel.architecture,and the resource scheduling strategy under service function non-integration.The experiment results indicate that our design shows the higher resource utilization and better performance.In the contrast experiments of resource utilization prediction,the BP neural network regression model achieves higher accuracy compared with the linear regression model.After optimizing system resource configuration with predicted values,the utilization of system resource is further effectively improved.