High-throughput Transmission In Data Center Network"Many-to-one"Traffic Mode

With the application and development of cloud computing technology in the fields of big data and artificial intelligence,the data center network as the core support platform of cloud computing has also been deeply developed in recent years.The data center integrates network resources,computing resources,storage resources,and other auxiliary resources(FPGA,ASIC,etc.)into a unified whole.It has become the infrastructure of computing-intensive business such as cloud computing,artificial intelligence,and mobile computing.The operational efficiency of data center network determines the service quality of these businesses.In the existing data center network,most applications still use TCP protocol for data transmission.And the characteristics of data center network,such as high-bandwidth,low-latency,are quite different from those of the traditional WAN.The operation efficiency of traditional TCP protocol in data center network is low,which will cause a series of problem,such as TCP Incast,TCP Outcast,etc.This paper focuses on the TCP Incast problem,which refers to the collapse of throughput was caused by buffer overflow of switch under the “many to one” data transmission mode in data center networks.The occurrence of TCP Incast will cause the network throughput drop sharply,and may cause the tasks to miss the deadline and be discarded,which will affect the quality of calculation results and user experience.Therefore,how to alleviate this problem to provide efficient data transmission services for cloud computing tasks such as distributed storage and web search is an important topic with practical application value.This paper starts from the characteristics of the data center network and the traditional TCP protocol,analyzes the root cause of the TCP Incast problem in detail and optimizes the efficiency of "many to one" traffic transmission mode of data center network from three aspects: single round data transmission,multi round data transmission and the impact of delayed ACK mechanism on heterogeneous data center network.The research contents and achievements are as follows:(1)We discussed the development of data center and the technologies of data center network,such as network topologies,software defined networks.Besides,we deeply studied the transmission protocols of data center network,summarized the main solutions to TCP Incast problem in recent years,and analyzed the problems of traditional TCP protocol running in the data center network and their root causes.(2)A TSL algorithm based on packet labeling for fast packet loss discovery and fast recovery is designed to solve the TCP Incast problem in a single round of data transmission.And a congestion adaptive window adjustment strategy is designed to solve the contradiction between the throughput and the number of concurrent support,which is caused by delayed ACK in heterogeneous data center environment.In view of the performance degradation of TCP and existing improved solutions under multiple rounds of data transmission,a window alignment strategy is proposed to dynamically adjust the initial value of the congestion window.(3)The above three strategies were implemented on the NS3 simulation platform,and the relevant simulation experiments were designed to verify its effectiveness and superiority.The experimental results prove that the solutions proposed in this paper can better adapt to the heterogeneous data center network environment and greatly improve the transmission efficiency of the “many to one” traffic transmission mode in the data center network.It is of great significance to improve the work efficiency of data center network equipment and provide efficient and reliable data transmission services for applications such as cloud computing and artificial intelligence.