Files in this item



application/pdfChi-Yao_Hong.pdf (3MB)
(no description provided)PDF


Title:Software defined transport
Author(s):Hong, Chi-Yao
Director of Research:Godfrey, P. Brighten; Caesar, Matthew
Doctoral Committee Chair(s):Godfrey, P. Brighten
Doctoral Committee Member(s):Caesar, Matthew; Nahrstedt, Klara; Gupta, Indranil; Feamster, Nick
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Data Center
Flow Scheduling
Transport Rate Control
Network Utilization
Software Defined Networking
Abstract:We advocate a software defined transport (SDT) architecture in which a transport controller schedules and dynamically re-schedules the flow sending rates based on current network conditions and the network operator's goals. This dissertation shows that this architecture provides both high flexibility, by allowing the operator to implement new transport policies as needed, and fine-grained flow control, by optimizing network resource allocation at flow-level in real time. We begin with proposing a fine-grained flow scheduling protocol to complete flows quickly and meet flow deadlines. Through extensive packet-level and flow-level simulation, we demonstrate that fine-grained flow control can significantly reduce mean flow completion times by 30% or more compared with TCP, RCP, and D3. We next design a software-driven controller which centrally allocates network resource such as bandwidth and routing paths for flexibility. In particular, we develop a prototype of our design for inter-datacenter wide area networks to achieve nearly optimal network utilization and service-level fairness. After that, we address network update problem to ensure bandwidth requirements during network updates subject to network capacity and switch memory constraints. Finally, we design and implement a fast, fine-grained flow-rate controller for data center networks. We show this design provides high scalability, by rate-controlling 95% of bytes of a cluster with several thousand servers within hundreds of milliseconds with a multi-threaded resource allocation algorithm, and application-level improvement, by reducing average shuffling times of MapReduce workload by 12-20%.
Issue Date:2015-01-21
Rights Information:Copyright 2014 Chi-Yao Hong
Date Available in IDEALS:2015-01-21
Date Deposited:2014-12

This item appears in the following Collection(s)

Item Statistics