Abstract: Because the traditional TCP protocol does not apply to the operating mode of data center, the TCP Incast problem occurs when there are common many-to-one traffic patterns in the data center, causing a visible throughput collapse of the application layer. Considering the characteristics of the data center, putting forward a comprehensive solution is the research objective of TCP Incast problem. This paper analyzes the root causes of the problem, enumerates the challenges of the problem, introduces the mathematical model based on the problem, analyzes and summarizes the recent solutions which are classified into link layer, transport layer and application layer, then from the effectiveness, deployment and other different aspects, makes a comprehensive comparison, finding that current solutions based on some specific points almost have drawbacks in different degree. Finally, this paper puts forward some feasible solutions to study the problem, and focuses on combining the technology of SDN and machine learning and designing a new transport protocol.

Key words: data center network, TCP Incast problem, throughput collapse

摘要： 目前传统TCP协议不适用于数据中心的工作模式，因此当数据中心中出现常见的多对一流量模式时会产生TCP Incast问题，造成应用层可见吞吐量崩溃。结合数据中心特点，提出全面的解决方案是解决TCP Incast问题的研究目标。围绕TCP Incast问题，深层次剖析了该问题发生的根源，简要概述了该问题面临的挑战，介绍了基于该问题所构建的数学模型；从链路层、传输层和应用层角度分析并总结了近十年具有代表性的解决方案，从有效性、可部署性等不同角度对所列举方案进行了全面对比，发现当前方案大都基于某个具体方面缓解该问题，均存在缺陷；最后提出了可行的解决该问题的研究方向，将关注点聚焦于SDN结合机器学习以及传输新协议。

关键词: 数据中心网络, TCP Incast问题, 吞吐量崩溃