Computer Science and Engineering, Department of

 

Date of this Version

Fall 12-1-2015

Citation

@phdthesis{ErtongZhang2015, title = {Bandwidth Estimation for Virtual Networks}, school = {University of Nebraska Lincoln}, author = {Zhang, Ertong}, year = {2015} }

Comments

A DISSERTATION Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfilment of Requirements For the Degree of Doctor of Philosophy, Major: Computer Science, Under the Supervision of Professor Lisong Xu. Lincoln, Nebraska: December, 2015

Copyright (c) 2015 Ertong Zhang

Abstract

Cloud computing is transforming a large part of IT industry, as evidenced by the increasing popularity of public cloud computingservices, such as Amazon Web Service, Google Cloud Platform, Microsoft Windows Azure, and Rackspace Public Cloud. Manycloud computing applications are bandwidth-intensive, and thus the network bandwidth information of clouds is important for theirusers to manage and troubleshoot the application performance.

The current bandwidth estimation methods originally developed for the traditional Internet, however, face great challenges in clouds dueto virtualization that is the main enabling technique of cloud computing. First, virtual machine scheduling, which is an importantcomponent of computer virtualization for processor sharing, interferes with packet time-stamping and thus corrupts the networkbandwidth information carried by the packet timestamps. Second, rate limiting, which is a basic building block of networkvirtualization for bandwidth sharing, shapes the network packets and thus complicates the bandwidth analysis of the packets. In this dissertation, we tackle the two virtualization challenges to design new bandwidth estimation methodologies for clouds. First, wedesign bandwidth estimation methods for networks with rate limiting, which is widely used in cloud networks. Bandwidth estimation fornetworks with token bucket shapers (i.e., a basic type of rate limiters) has been studied before, and the conclusion is that “bothcapacity and available bandwidth measurement are challenging because of the dichotomy between the raw link bandwidth and thetoken bucket rate”. Our methods are based on in-depth analysis of the multi-modal distributions of measured bandwidths. Second, we expand the design space of bandwidth estimation methods to challenging but not rare networks where accurate and correctpacket time information are hard to obtain, such as in cloud networks with heavy virtual machine scheduling. Specifically, we designand develop a fundamentally new class of sequence-based bandwidth comparison methods that relatively compare the bandwidthinformation of multiple paths instead of accurately estimating the bandwidth information of a single path. By doing so, our methods useonly packet sequence information but not packet time information, and are fundamentally different from the current bandwidthestimation methods that all use packet time information. Furthermore, we design and develop a new class of sequence-basedbandwidth estimation methods by conveying the time information in the packet sequence. Sequence-based bandwidth estimationmethods estimate the bandwidth information of a path using the time information conveyed in the packet sequence from another path.

Adviser: Professor Lisong Xu

Share

COinS