An Effective Method for Mitigating TCP Incast Throughput-collapse Problem of Data Center Networks

• Main Authors: Wei-Ru Tseng, 曾韋儒
• Other Authors: 林偉
• Format: Others
• Language: zh-TW
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/23640995996321846532

碩士 === 國立中興大學 === 資訊網路多媒體研究所 === 100 === With advances in cloud computing technology, a growing number of data centers have been constructed in the past few years. Networks play an important role in determining the performance of data centers. Issues relevant to network performance of data centers have attracted a great deal of attention in research. At present, many data centers simply employ TCP/IP as well as Ethernet for constructing the networks, taking advantage of their cost-effectiveness and ease of use. However, when engaging many-to-one transmissions, networks using TCP are prone to collapse in throughput. This is due to the drastic reduction in effective bandwidth when the multiple senders attempt to transfer packets to the same receiver. This is called the TCP Incast problem. The Incast problem leads to a serious degradation in network performance, and as a consequence data centers fail to deliver the QoS they guarantee. There is a class of applications, such as cluster storage, MapReduce and web search, that need to process numerous data requests from various servers. This happens to fit the communication pattern leading to the Incast problem. Mitigating the Incast problem, therefore, can significantly improve the network performance for these applications on data centers. The TCP Incast problem has been extensively studied over the past few years. But there is no consensus of a cost-effective approach to solving the problem. Many of the proposed methods so far are of no practical use, while others are too costly to implement. These proposed methods tend to ignore the effect of network topology and characteristics. In this thesis, we propose to investigate the effect of network topology and characteristics on the performance of Incast. We intend to tackle the Incast problem through the inclusion of network topology and characteristics. The Incast problem is divided into two subproblems at two different levels of data center networks. At the lower network level, we aim to reduce packet loss rate by resizing packets and increasing buffer capacity. We make use of interconnection network in the Hyper-Fat-tree Network (HFN) topology to break up the potential hot spots in the presence of overwhelming requests, therefore, reducing the occurrences of Incast bottleneck. At the upper network level, we dynamically vary the length of server request unit under various loading situations. In this manner, the effective bandwidth is improved and the Incast problem is alleviated. We simulate our proposed method based on the NS-2 simulator. The simulation results show that the proposed method can effectively mitigate the Incast problem as well as eliminate the possibility of throughput collapse. For example, in the case of 3600 servers, the proposed method can achieve a 90% effective bandwidth without suffering the Incast problem. This allows users to efficiently access data in the data center, and the quality of service is improved.