Modeling And Optimization Of The Irregular Multi-Level Fat Tree Networks

The fat tree network is a multi-level interconnection network with the advantages of good scalability,high bandwidth,no deadlock and simple routing algorithm.It is widely used in multi-level switching networks,on-chip networks,high performance computing,data centers and other fields.A variety of applications and large data,cloud computing and other new businesses require the network with the fast parallel processing capability,and people continue to reform the traditional fat tree network to improve its performance.However,the cost problem has become increasingly prominent,and gradually becomes an important factor which can not be ignored and the study focus in the fat tree networks.Sponsored by the sub-project,the research and industrialization of the key technologies of Tens Terabit Packet Transport Network,from the Ministry of Industry and Information Technology,this dissertation aims to study the improvement and optimization of the fat tree network topology.This dissertation gives a brief review of the characteristics and development of tree-shaped networks(TSNs),and summarizes the TSNs metrics,and analyzes the influence of each parameter on the TSNs performance.It introduces the TSNs application fields and presents a brief summarization of the TSNs research status.This dissertation summarizes the improvement of TSNs topology and analyzes the factors that affect the cost of a switching module.The methods and techniques used to reduce the network cost are described and summarized.The main contributions in this dissertation include three following parts:(1)To reduce the cost of traditional fat tree networks,this dissertation made an improvement to the traditional fat tree network topology.An irregular multi-level fat-tree network(IMLFTN)is proposed and its analysis model is established.The network node uses a special basic switching module(BSM)in which the number of upward ports is fixed while the number of downward ports keeps variable.The BSMs used in the first level are chosen according to the total number of users accessing the network,and the BSMs locating at other levels are accordingly determined and connected together to form an IMLFTN.(2)This dissertation introduces the exchange mechanism in the network and analyzes the traffic distribution of each switching module in the IMLFTN,and fits the cost function of a switching module.The cost function considers the numbers of upward and downward ports,and the traffic passing through it.When the number of users accessing the network is given,we could use the fitted cost function to calculate the total cost of all the networks constructed by the combination of switching modules with different number of ports.The IMLFTN with optimal total cost is selected as the optimization structure.(3)The instance of optimal network structure under different users in this dissertation is implemented via Java programming,and a comparison of the calculation results to the cost of traditional fat tree network is given to verify the feasibility and effectiveness of the proposed model.It also analyzes the effect of the relevant parameters on the overall network cost.The results of this dissertation show that the proposed irregular multi-level fat tree network has a large advantage in cost when constructing the network with the same number of users.The research in this dissertation is expected to guide the construction of actual network architecture and optimize the network cost.Combined with the users that actually access into the network,we can design a network with certain scalability and optimal cost in a certain range of predictable growth.