数据中心网络中拥塞现象和不公平现象的分析和改进

发布时间：2019-03-11 15:11

【摘要】：随着数据中心应用提出的需求不断变高，数据中心网络作为数据中心的重要组件，其负载也变的越来越大。这样导致拥塞情况频繁的发生，容易形成一种被称为incast的吞吐量大幅下降的现象。而数据中心应用程序的流量特征又会导致出现一种被称为outcast的不公平现象。通过深入了解incast现象的几个关键参数，可以发现，RTO min在其中起着至关重要的作用。随着发送方数量的增大，网络拥塞程度不断提高，这样会导致丢包和超时现象出现。严重的拥塞导致发送方无法通过收到三个冗余ACK来进入快速恢复状态，只能等待超时，从而影响网络吞吐量。在众多解决incast问题的算法中，传输层算法属于一类较好的算法，因为此类算法对网络基础设施的要求和改动较小，同时易于实现。所以在本文中，通过设计一种基于ACK回复变化率的传输层协议，较好的解决了incast问题。协议利用ACK回复变化率和对理论最大拥塞窗口的估计调节当前拥塞窗口，有效处理了incast现象中吞吐量下降的问题。数据中心的数据流还存在一种不公平现象，即P. Prakash发现的被称为outcast的现象。这种现象的表现就是RTT小的流，其吞吐量小于RTT大的流，这与传统TCP协议遵循的RTT与吞吐量成反比的准则完全相反。P. Prakash在验证了outcast现象的广泛存在后，给出了一个基于端口阻塞的解释。但是经过本文仔细的分析后，我们给出了outcast现象的实质原因。即RTT不同的流在物理链路上的分布不均，和数据中心上层应用的特征所导致的RTT不同的流在拥塞窗口大小上的差异。根据该现象的本质原因，本文建立了相应的吞吐量数学模型。最后，通过设计一种基于窗口通知的协议，，解决了outcast现象。协议通过测量拥塞窗口的平均值，统一了RTT不同的流在当前数据块发送结束时的拥塞窗口大小，使得RTT小的流的吞吐量得到提高。将以上几个算法在ns-2仿真平台中进行实验后，实验结果证明outcast现象的本质原因的正确性和两种解决算法的有效性。
[Abstract]:With the increasing demand of data center application, the load of data center network, as an important component of data center, becomes more and more heavy. This leads to the frequent occurrence of congestion, and it is easy to form a kind of phenomenon, called incast, which has a large decline in throughput. The traffic characteristics of data center applications lead to an unfair phenomenon called outcast. Through in-depth understanding of several key parameters of incast phenomenon, it can be found that, RTO min plays an important role in it. With the increase of the number of senders, the degree of network congestion is increasing, which will lead to packet loss and timeout. Due to the serious congestion, the sender cannot enter the fast recovery state by receiving three redundant ACK, and can only wait for the timeout, thus affecting the throughput of the network. Among the many algorithms to solve the incast problem, the transport layer algorithm belongs to a better class of algorithms, because the network infrastructure requirements and changes of these algorithms are small and easy to implement at the same time. In this paper, a transport layer protocol based on ACK recovery rate is designed to solve the incast problem. The protocol adjusts the current congestion window by using the rate of change of ACK recovery and the estimation of the theoretical maximum congestion window, and effectively deals with the problem of throughput decline in the incast phenomenon. There is also a kind of unfair phenomenon in the data center data stream, that is, the phenomenon called outcast discovered by P. Prakash. The performance of this phenomenon is the small flow of RTT, whose throughput is smaller than that of large RTT, which is completely contrary to the principle that RTT and throughput are inversely proportional to the traditional TCP protocol. After verifying the widespread existence of outcast phenomenon,. P. Prakash has verified the widespread existence of outcast phenomenon. An explanation based on port blocking is given. However, after careful analysis in this paper, we give the real reason of outcast phenomenon. That is, the distribution of different RTT flows on the physical link is uneven, and the difference of the congestion window size between the different flows of RTT caused by the characteristics of the upper application of the data center and the distribution of different flows on the physical links. According to the essential reason of this phenomenon, the corresponding mathematical model of throughput is established in this paper. Finally, a protocol based on window notification is designed to solve the problem of outcast. By measuring the average value of the congestion window, the protocol unifies the congestion window size of different RTT flows at the end of the current block transmission, which improves the throughput of the small RTT stream. The experimental results show that the essential reason of the outcast phenomenon and the effectiveness of the two algorithms are correct after the above algorithms are tested on the ns-2 simulation platform.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP308;TP393.06

【共引文献】