中文摘要：

分析了数据中心的应用特性、流量特征以及目前TCP协议存在的不足，指出目前数据中心的大规模交互式网络应用具有软实时性，因而受时限约束，而网络应用的划分．聚合设计模型及采用的公平共享的拥塞控制协议，是导致网络流错过时限的主要原因。据此，提出了一种基于优先级的时限感知的数据中心传输控制协议（PD^2TCP），一种新的网络拥塞控制算法。该算法在交换机端，根据瞬时队列长度和单一门限进行显式拥塞通告（ECN）标记；在主机端，根据流的时效性需求及其历史信息，赋予其不同的优先级，并根据流的优先级和网络的拥塞程度调整拥塞窗口。同时在小规模的真实环境中和大规模的仿真环境中对PD^2TCP的性能进行评价。实验表明，与时限感知的数据中心TCP（D^2TCP）相比，PD^2TCP错过时限流的比例降低了65％，流完成时间的99^th分位数降低了45％，并且几乎没有降低延迟不敏感的背景流的吞吐率。PD^2TCP能够和TCP共存，因而可以在真实环境中部署。

英文摘要：

The characteristics of datacenter applications as well as their traffic flows and the the inefficiency of state-of- the-art transmission control protocol （TCP） are analyzed, and it is pointed out that the large-scale interactive web applications in today＇ s datacenters have soft-real-time natures so there are deadline constraints associated with network flows, and that misseddeadlines are mainly caused by the partition-aggregate design pattern of the applications, combined with the fair-sharing congestion control protocols. Based on the analysis, the Priority-based Deadline-aware Datacenter TCP （PD^2TCP）, a novel congestion control algorithm is proposed. PD2TCP assigns different priorities to flows according to their timeliness requirements and history information, and modulates the congestion window based on the priorities and the extent of congestion. The performance of the PD^2TCP was evaluated in a small-scale real environment and a large-scale environment, and the results showed that, compared to the deadlineaware datacenter TCP （D^2TCP）, the PD^2TCP reduced the fraction of deadline-missing flows and the 99^th percentile normalized flow completion time by 65% and 45%, respectively. The pD^2TCP can co-exist with TCP, thus can be deployed in real datacenters.