On The Optimization Of Routing And Transport Protocol In Data Center Networks

With the rise and boom of big data, cloud computing and high performance computing, data center, the infrastructure that hosts such applications, has become the industrial and academic hotspot. The feature of data center network (DCN) is much different from that of conventional wide area network (WAN), and the demand of DCN-hosted application and the traffic patten vary tremendously, which bring up the new challenge on availibility and reliability. In this thesis, we research on the performance of high throughput and low latency of DCN, and explore the academic field of routing and flow control mechanism.The first topic is load-balancing routing in data center. With the feature of full bisection bandwidth, Fat-tree topology is widely adopted in data center networks. However, the state-of-the-art load balancing routing selection function works inconsistently in various traffic patten, and the implementation is complex. Based on the insight of load balance, we adopt the configu-ration of predetermined and periodic sequential connection of servers and top-level switch. This selection function spread the traffic evenly on all the top-level switch for load balance, and it has no prefence on traffic patten. Moreover, there is no need for upward switch buffer or arbitrator, and flow control mechanism is simplied to single stage, which reduce the complexity of hardware implementation and control plane message.The second topic is packet loss induced by bursty concurrent traffic. The synchornized traffic overflows the shallow-bufferd aggregator switch and induce packet discard, thus leading to time-out retransmission and throughput decrease. DCTCP(Data Center TCP) can address this problem to a certain extent, but is limited by the ceiling bound of concurrent senders. Based on DCTCP, we revise ECN protocol to stamp a specific mark on packet that is supposed to be discarded, and route it back to sender as explicit notification of packet loss. The sender retransmit is after carrying on the flow-based binary exponential backoff algorithm. The enhanced DCTCP protocol achieves high throughput in cost of silghtly raised retransmission ratio, and eliminates timeout under multiple scenario.