Research On Optimization Of Congestion Control Algorithms In Data Center

The Transmission Control Protocol (TCP) is widely used in the transport layer. After decades of development, the TCP protocol has spawned dozens of versions, and can be suitable for the high speed network and wireless network scenarios. However, in the data center network with high bandwidth and low latency, TCP faces new challenge. In the data center, many to one transmission mode causes Incast problem in TCP layer. TCP Incast refers to the TCP protocol throughput collapse problem with the increase in the number of concurrent senders. Incast problem is first reported by D. Nagle in the design of distributed storage architecture. Further research has found that this problem widely exists in partition-aggregation transmission mode of data center and has attracted many scholars’ attention.Aiming at the Incast problem in the data center of the network transmission, this paper has conducted in-depth analysis. According to the experiment results of the authoritative network simulator (NS-2), we analyzed the effect of different parameters on Incast. The results show that, to increase the cache can to some extent alleviate the Incast problem, but arouse other problems, such as big queueing latency. In addition, the constraint of the sending window and the decrease of RTO can also improve the throughput of concurrent transmission. Then we analysis the Data Center TCP algorithm and its backwards. According to the effect of reducing the minimum retransmission timeout (RTO), we propose the dynamic RTO algorithm which tunes the RTO value according to the congestion level. Through the NS-2 simulation platform, we prove that the dynamic RTO algorithm can improve the number of concurrent senders in the premise of maintaining the throughput. In the workflow of online service system, in order to improve the quality of service, developers add soft real-time constraints to data transmission. For the soft real-time constraints on deadline flows, we propose a deadline flow scheduling algorithm to improve the efficiency of data transmission of deadline flow. The simulation results show that this algorithm can improve the completion rate of deadline flow by ten percent.In this paper, we aim to improve the congestion control algorithm of data center, to alleviate the Incast problem, and to improve the quality of service of the online business. Based on the DCTCP algorithm, we propose dynamic RTO algorithm and deadline flow scheduling algorithm. These algorithms can improve concurrent transmission throughput and the on-time completed transmission ratio of deadline flow, effectively improving the transmission performance of the data center network.