数据中心网络的流量控制: 研究现状与趋势

As a strong foundation for the rapid storage and efficient processing of massive data, the data center has become a hot spot in academia and industry in recent years. Traditional TCP is difficult to meet the demand for data center transmission in high throughput, low latency and loss-free aspects. Based on the comparison between the traditional TCP design target and the transmission target in the data center network, this paper summarizes the research status of data center traffic control. Traffic control refers to the control of traffic rates and sending rules. Therefore, this paper introduces the traffic control technology based on congestion control and traffic engineering and makes a comparative analysis of the above technology from the aspects of control mechanism, expansibility and technical feasibility. Finally, this paper summarizes and looks forward to the future research trend of data center traffic control technology. According to the existing researches, we find that: (1) Considering the cost and performance, the most suitable traffic control algorithm for TCP/IP data center is DCTCP, and the most suitable traffic control algorithm for RDMA data center is DCQCN. Other researches require expensive custom hardware, which is difficult to deploy. (2) The traffic control technology is a technology of fair utilization of limited resources. Therefore, the performance of the technology can be improved by acquiring more relevant information or exchanging with other resources. E.g. ECN, RTT, traffic size, flow deadline. (3) Among the three main research points of congestion control, flow scheduling and load balancing, the mainstream algorithm system only focuses on one or two of them. (4) Smart NICs and programmable switches are widely used in the research of the data center network. The programmability of smart devices can bring new features to new technologies. In the end, the research directions are prospected. (1) A unified flow control test platform is needed. Different algorithms use different test environments, so it is difficult to evaluate them together. (2) Congestion control, flow scheduling, and load balancing studies need to be considered together. (3) Traditional distributed traffic control cannot be accurately scheduled due to insufficient information. As data centers grow in size, centralized controllers become network bottlenecks. The tradeoff between centralized and distributed control requires careful consideration. (4) High-performance programmable smart devices need to be developed and deployed. RDMA has become a hot topic in industry and academia. At the same time, programmable network devices greatly enhance the flexibility and rapid deployment of the network. (5) Traffic control design for specific application scenarios. The performance of the algorithm is improved by acquiring more relevant information or by exchanging other related resources. More resources are available in specific application scenarios. (6) With its strong self-adaptability and self-learning ability, artificial intelligence provides a set of effective decision-making tools for various research fields. The combination of artificial intelligence technology and network transmission technology is also a hot topic in the future. In summary, with the in-depth study of the data center, traffic control will become the most important basic performance tool for the data center, especially for the future high throughput, low latency requirements.