A layer 2 multipath fabric using a centralized controller

• Author(s): Júlio, Fábio José Correia
• Date: 2013
• Persistent ID: http://hdl.handle.net/10362/11139
• Origin: Repositório Institucional da UNL
• Subject(s): Data center; Routing; OpenFlow; Policies; Fault tolerance
• Document Type: Master thesis
• Language: English
• Advisor(s): Amaral, Pedro
• Contributor(s): RUN

Description

Dissertação para obtenção do Grau de Mestre em Engenharia Electrotécnica e de Computadores

Ethernet is the most used L2 protocol in modern datacenters networks. These networks serve many times like the underlying infrastructure for highly virtualised cloud computing services. To support such services the underlying network needs to be prepared to support host mobility and multi-tenant isolation for a high number of hosts while using the available bandwidth e ciently and maintaing the inherent costs low. These important properties are not ensured by Ethernet protocols. The bandwidth is always wasted because the spanning tree protocol is used to calculate paths. Also, the scalability can be an issue because the MAC learning process is based in frame ooding. On layer 3 some of this problems can be solved, but layer 3 is harder to con gure, poses di culties in host mobility and is more expensive. Recent e orts try to bring the advantages of layer 3 to layer 2. Most of them are based in some form of Equal-Cost Multipath (ECMP) to calculate paths on data center network. The solution proposed on this document uses a di erent approach. Paths are calculated using a non-ECMP policy based control-plane that is implemented in an OpenFlow controller. OpenFlow is a new protocol developed to help researchers test their new discovers on real networks without messing with the real tra c. To do that OpenFlow has to be supported by the network's switches. The communication between systems is done by SSL and all switches features are available to the controller. The non-ECMP policy based algorithm is a di erent way to do routing. Instead of using unitary metrics on each link, one policy is chosen for each link. The use of policies opens the possibility to consider very di erent paths as having the same forwarding preference increasing the number of used paths. Our approach uses the recent Backbone Provider Bridging (PBB) standard that adds extra header information to the Ethernet frame and provides isolation between customer and network address space improving scalability.