Fat-tree数据中心网络中多播数据传输恢复

发布时间：2018-04-23 05:17

本文选题：数据中心 + 超额认购　；参考：《西南大学》2017年硕士论文

【摘要】：传统的数据中心大多采用一种被称为Fat-tree的多根树模型,这种模型利用丰富的多重路径传输高对分带宽,在它的根部部署了很多的核心交换机去实现功能和规模的扩展。Fat-tree数据中心网络可以在任何两个主机间提供多重的链路连接,进而提供大的聚合带宽,然而在Fat-tree中不均匀的流量分布可能引起低效率的传输,使系统的网络带宽不能得到充分的利用,这种情况也引起了网络处理能力的衰减,导致了系统性能的不可预测。这种情况阻止了系统充分发挥并行链路的高性能特性,并且会引起很多不可预测的问题。为了解决以上的挑战,本文提出一种数据中心Fat-tree结构中的多播数据修复策略(MRDC)。对于多播树中的每一个接收者,MRDC给它分配一个数据备份系统,系统根据服务器所处的位置进行数据的恢复。因此,当多播传输完成的时候,数据备份系统可以侦测到因服务器接收数据失败而导致的数据包丢失;此时,与接收数据失败的服务器跳数最少的邻居服务器将向接收失败的服务器传输数据包,直到接收成功。此外,在传输数据的过程中,MRDC将向多播管理者通报所有的接收服务器的信息,管理者将选择最优的数据恢复链路,进而减少在数据恢复过程中可能产生的链路拥塞。另外,数据中心中的多播路由算法通过调整传输速率和接收速率可以有效地避免拥塞现象的出现。在数据中心网络中,两个设备之间有许多条链路连接,也就是说从一台设备发送数据到另外一台设备有多条路径可以选择,如果在大规模的数据传输的时候不对链路加以选择很容易引起超额认购现象的发生,这种情况会造成数据传输的失败,影响系统的性能,造成许多不可预测的现象,因此本文设计了一种链路状态的计算算法,可以根据链路的历史状态和现行流量负载计算出它的超额认购概率,进而筛选出最优的链路,优化系统的整体性能。因此它很大程度上减少了带宽的消费和传输的延迟,提高了网络吞吐量。
[Abstract]:Traditional data centers mostly use a multi-root tree model called Fat-tree, which utilizes rich multipath transmission with high bandwidth. There are many core switches deployed at its root to enable functional and scalable. Fat-tree data center networks that provide multiple link connections between any two hosts, thus providing large aggregate bandwidth. However, the uneven traffic distribution in Fat-tree may lead to inefficient transmission, which makes the bandwidth of the system not be fully utilized, which also leads to the attenuation of the network processing capacity and the unpredictable performance of the system. This situation prevents the system from making full use of the high performance characteristics of parallel links and can cause many unpredictable problems. In order to solve the above challenges, this paper proposes a multicast data repair strategy in data center Fat-tree architecture. For each receiver in the multicast tree, MRDC assigns it a data backup system, which recovers the data according to the location of the server. Thus, when multicast transmission is complete, the data backup system can detect packet loss caused by the server's failure to receive data; at this time, The neighbor server that has the least number of hops with the server that failed to receive the data will transmit packets to the failed server until it is received successfully. In addition, during the transmission of data, the MRDC will inform the multicast manager of all the information of the receiving server, and the manager will select the optimal data recovery link, thereby reducing the possible link congestion during the data recovery process. In addition, multicast routing algorithm in data center can effectively avoid congestion by adjusting transmission rate and receiving rate. In a data center network, there are many links between two devices, that is, there are multiple paths to choose from one device to another. If the link is not selected during the large-scale data transmission, it is easy to cause oversubscription, which will result in the failure of the data transmission, affect the performance of the system, and cause many unpredictable phenomena. Therefore, this paper designs a link state calculation algorithm, which can calculate the oversubscription probability according to the historical state of the link and the current traffic load, and then select the optimal link and optimize the overall performance of the system. Therefore, it greatly reduces bandwidth consumption and transmission delay, and improves network throughput.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.0

【参考文献】