The Study And Implementation Of Time Series Based Bandwidth Allocation Strategy In Cloud Datacenter

Virtualization technology has made great progress in recent years and promoted the development of cloud datacenters. Cloud datacenter provides satisfying services for tenants to configure memory, storage, CPU for their virtual machines. However, network is the resources for sharing and competing, it is difficult to balance cost of service providers and SLA of tenants. Most cloud datacenters apply tree topology, network utilization is quite low in the access layer. Service providers raise the oversubscription ratio to reduce the cost, however, it makes bandwidth competition more incentive.Starting from network architecture of cloud datacenter, this paper firstly study the traffic characteristics of virtual machines macroscopically. Secondly, a prediction model is proposed to forecast load state of bandwidth according to bandwidth usage of each VM. Finally, allocation strategies are formulated based on prediction and bandwidth resource is allocated to each VM in advance.The key points of the research are attached as following:1. Investigate the existing cloud platform and conclude the solutions of network architecture, analyze the influence of oversubscription and study the characteristics of network traffic, in order to provide sufficient theoretical basis and application scenarios for the research.2. Design a centralized network controller based on OpenFlow protocol to collect data of bandwidth usage, build prediction model based on time series analysis to forecast load state of bandwidth in the future.3. Propose bandwidth allocation strategy based on prediction of load state of bandwidth and divide these states in three categories which are low, high and full utilization, each has a corresponding method to allocate bandwidth.4. Implement the allocation system based on allocation model and OpenStack. The experiment replay the real-world traffic in the virtual subnet and compare the results with static allocation, which is the mostly applied method.In conclusion, our research fully consider the dynamic bandwidth demands of different VMs, the resource is allocated in advance based on prediction of bandwidth load state. The results show that both bandwidth utilization and total traffic are improved in the access layer.