随着云业务的迅猛发展，其背后支撑的物理基础设施数据中心吸引了工业界和学术界的广泛关注。当前大部分数据中心采用水平扩展的方式连接大量廉价的交换设备和服务器来支持大规模数据存储和计算，服务器之间需要频繁的通信来协作完成任务，所以数据中心网络通信延时成为影响用户体验的一个重要因素。而由于数据中心网络在拓扑、流量等方面独特的特性，传统网络中工作良好的传输控制协议在数据中心网络中遇到了很多挑战，显著增加了业务完成时间。本课题首先提出一种集散式传输控制协议框架，并基于将网内缓冲池从网络容量中去除的基本理念设计了零排队延时的拥塞窗口分配算法；然后提出基于漏桶算法的包延迟机制解决高并发流模式下的网络拥塞问题；在此基础上将拥塞窗口算法根据流截止时间进行扩展，保障流级别的完成时间需求；最后，通过集中式确定任务的全局优先级标识，并配合分布式的流延时保障机制，减小任务完成时间。

With the proliferation of cloud services, data center, the physical infrastructure of supporting cloud computing, attracts much attention from both industry and academia. Most of current data centers employ the scale-out model to connect a large number of cheap switches and servers and thus support large-scale data storage and computing. Frequent communication between servers is required to concurrently complete a job. Thus, the performance of data center networking poses a great impact on the user experience of cloud services. However, due to the special characteristics of data center networks in terms of topology, traffic and so on, traditional transport control protocol faces great challenges and causes severe performance degradation of cloud services. This proposal firstly proposes a semi-centralized transport control protocol framework, and presents a zero-queuing congestion control mechanism by excluding the in-network buffers from the network capacity. Then a packet delay algorithm is proposed to avoid network congestion with high concurrent flows based on the leaky bucket algorithm. Subsequently, the congestion control algorithm is extended according to flow deadlines to guarantee flow-level deadline requirements. Finally, task completion time is reduced by centrally determining the task global prioritization ID and cooperating with the distributed flow-level deadline guarantee mechanism.