Research And Design On Rate Control In Data Center Network

Nowadays,data centers(DC)have become the underlying infrastructures for computing,storage and data transmission,and are hosting an increasing number of applications.These applications vary in network requirement,which brings new challenges for data center networks(DCNs);it not only requires DCN to fully complete the data transmission,but also requires a fast,stable and reliable data transmission.Hence,rate control is crucial in DCNs;rate control is efficient in mitigate the network congestion in DCNs,and thus improve the effectiveness of date transmission.However,current methods on rate control are insufficient,and face two challenges.First,current methods cannot guarantee the bandwidth to be allocated to applications in an "on demand,dynamic,and fair" manner.Second,in terms of the congestion control in network transport layer,current methods are unaware of the application demand,making it difficult to adaptively control the congestion window.In this paper,we conduct the rate control research at both application layer and network transport layer.First,at the application layer,we propose an application demand based Borrow-Return-Preemptive(BRP)algorithm,which works in a dynamic way to improve the serving rate of application requests and network utilization.If the bandwidth is reserved too much,the application will lend the idle bandwidth to other applications.Otherwise,the application will jointly combine the algorithms of bandwidth borrowing,bandwidth returning,and bandwidth preempting to meet application requirement.Second,at the transport layer,we design a kind of application-aware congestion control scheme,a basic scale model to reduce the average completion time of applications,which coordinately consider the congestion degree,the urgency of the application and the priority of the traffic flow.As the congestion degree become larger,the congestion window will become smaller.The urgency and higher priority flow is larger than the infrequent and low priority flow on the congested window.Experiments show that compared with the fair allocation method,our BRP method improve bandwidth utilization by up to 19.7% and serving rate of application requests by up to 37%,with an average increase of 11.8% and 21.9%,respectively.Through theoretical argumentation and experiment analysis,our basic scale model has the ability to meet the requirements of different application priorities and urgencies.