Modeling And Optimization Of A Special Tree-shaped Network

Interconnection networks(INs) are widely used in the systems of high-performance computers, data center networks and multistage switching networks. The ever increasing complexity of these systems results in the increasing importance of INs, and enabling INs to be one of the key factors which determine the overall performance of a system. Tree-shaped network is an IN that is now widely used. It has the advantages of topology scalability, nondeadlock and simplicity in routing algorithms. When constructing and optimizing a treeshaped network, we could use queueing network model to select the optimal network parameters and obtain the optimal network architecture designed according to the actual needs.The research in this dissertation is based on a special tree-shaped network(STN), which is proposed by our supervisor's research group. It is constructed by connecting the basic switch modules(BSMs), and each BSM has multiple low-rate ports and a high-rate port, which can reduce the comprehensive cost of a switch element. After a brief description of the topological structure of the special tree-shaped network and its routing algorithm, we construct the equivalent queueing node for a single BSM by analyzing the procedure of forwarding packets within each BSM. Then, we can get the queueing network model when connecting these queueing nodes of all BSMs in accordance with the interconnected relationship of BSMs in original network. We use the method of decomposition to analyze the performance of the network based on the queuing network that is constructed. The network is divided into downward and upward sub-networks. The performance of each subnetwork could be obtained according to the order of packets transmission, and therefore we get the performance of the network by processing the performance outcome of each subnetwork based on the original traffic model and the routing probability of a packet. We have implemented the design and programming of the simulation programs according to the running process of such a queuing network, obtaining the network performance parameters and verifying the theoretical analysis. In addition, we analyze the effects of offered traffic, cache buffer size, network topology and caches allocation schemes on the network performance, and the conclusions are outlined as follows:(1) The influence of the offered traffic on network performance is mainly confined to the same network architecture. For different network architectures, if the offered traffic increases, the throughput, latency and loss rate will increase in different degrees, respectively. When the number of user ports or terminals is certain, the performance gap between different network architectures becomes larger as the offered traffic increases. However, the variation of the offered traffic does not affect the choice of the optimal network architecture.(2) We find that the network performance, such as throughput, latency and loss ratio, is closely related to the needs of actual traffic, buffers of BSMs, topological structure and caches allocation schemes. Therefore, we could get the optimal network architecture by carefully selecting the optimal network parameters based on the actual needs of traffic.The research in this dissertation is expected to be used in the parameters selection and optimization of actual network architecture. It means that if we can select topology structure and set caches within each BSM reasonably according to the actual traffic requirements, and the optimal network could be constructed.