【作者】 赵彤；

【导师】 王翠荣；

【作者基本信息】 东北大学 ， 计算机技术（专业学位）， 2012， 硕士

【摘要】 最近几年,随着网络技术的不断进步,现代数据中心进入飞速发展阶段。现代数据中心为不同企事业单位提供各式各样的并发服务,承载着大量的交互式数据,需要更高性能的网络传输。但实际上由于多方面原因,尤其是应用程序开发人员不深入了解TCP协议机制,编写的应用程序可能导致主机负载过重、网络拥塞、缓冲区溢出等,使得运行于现代数据中心的应用程序往往经历了低吞吐量和高延迟。另外TCP协议的Nagle等算法也可能导致应用程序发送和接收数据缓慢,传输性能降低。因此查找影响现代数据中心传输性能下降的原因和异常节点定位成为本文要解决的主要问题。现有的基于SNMP协议的网管软件虽然也能检测网络的吞吐量、延迟抖动等性能参数,但不能够细粒度地准确定位是哪个具体应用程序或者网络节点引起。本文针对现代数据中心存在大量TCP流量,根据TCP数据包和Socket日志,深入分析了应用程序端到端的重传、超时等参数传输状态,提出了基于皮尔森相似度的异常节点查找算法,使用遗传算法进一步分析并找到了通过异常节点的所有TCP连接中引起该节点性能下降最严重的连接。根据上述分析和给出的算法,本文设计并实现了一种基于TCP的数据中心网络性能监测系统。最后,分别在实验室和学校数据中心进行了实验验证,通过人为制造一些异常,验证了以上算法的准确性。在实验室搭建了一个数据中心网络实验环境,通过使用Libpcap抓包库被动采集网络数据包,最终实现了查找性能异常节点的目的。在学校数据中心的每个服务器上安装了本人设计的网络监测软件,对采集到的实时数据包和TCP的Socket日志进行分析,证明该性能监测系统可以准确地找出异常节点与应用程序。更多还原

【Abstract】 With the development of network technology in recent years, data center enter a rapid development stage. Nowadays data center provides a wide range of concurrent services for various enterprises and institutions. Because of data center carries a large number of interactive data, it needs better performance of network transmission. But in fact, some application developers do not understand the TCP protocol mechanisms deeply. Applications written by them may lead to a host some problems like overload, network congestion, buffer overflow. And these problems will make the applications which run on the data center experience a low throughput and high latency. In addition, the Nagle algorithm of TCP protocol may also lead applications to send and receive data slowly. So applications will have a lower transmission performance. All in all, find out the reason of the influence transmission performance of data center and locate the abnormal nodes becomes the main problem to be solved in this paper.Although some existing SNMP-based network management software can detect the network performance, they can’t find out the accurately location of the specific application or network node which make that so. In this paper, after having known that there are lots of TCP flows in data center, by making statistic analysis of the TCP packets and Socket logs, we deeply analysis and record applications’ end-to-end parameters’ state like time-out retransmissions. This paper proposes an algorithm based on Pearson correlation coefficient to find the abnormal nodes which decline the transmission performance of data center. Then by using genetic algorithm, we can analysis the abnormal nodes and find the worst several end-to-end connections on each abnormal node. According the analysis and algorithms mentioned above, this paper designs and achieve a date center network performance monitor system based on TCP.Finally, this paper did some experiments in the laboratory and school data center, and verified the accuracy of the above algorithm by making abnormal situation. In the laboratory a data center network environment was set up. And it achieved the purpose of searching abnormal nodes by collecting network packets with Libpcap. In the school data center, the designed network monitoring software was installed on each server. By analyzing the collected real-time data packets and the TCP Socket logs, this paper proved that the monitoring system could accurately identify abnormal nodes and applications.更多还原