With respect to the evolution of its architecture, the Internet has been a victim of its own success. While the Internet has driven remarkable innovation in the applications which run over it, its rapid growth in size and complexity has introduced critical looming challenges and problems. One result of this growing complexity is referred to as “Internet Ossification” – namely, an inability to experiment with, evolve, or redesign the underlying protocols and mechanisms comprising the Internet's architecture. Moreover, as the resource and performance demands placed by users upon today's Internet continue to increase, the network's predisposition against experimentation and evolution makes finding (and more importantly, deploying) solutions to these demands highly challenging to say the least. In recent years, a number of approaches have been undertaken toward attacking the ossification problem on different fronts. One particular approach to this problem is known as “Software Defined Networking” (SDN), whereby evolvability is introduced into the underlying Internet infrastructure through flexible, programmable networking hardware with standardized interfaces. Numerous architectures to realize SDNs have been proposed, and most notably, the OpenFlow initiative has been gaining traction and widening deployment in campus networks around the United States, Europe, and Asia. The programmability introduced by SDN offers an appealing mechanism toward enabling experimentation, virtualization, and deployment of new protocols and network architectures within today's networks. While the introduction of programmability into the network architecture creates significant potential for new approaches in network design, it also raises significant questions about the scalability and limitations of these approaches.
|Author||Wundsam, Andreas and Levin, Dan and Seetharaman, Srini and Feldmann, Anja|
|Title of Book||Proceedings of Usenix Annual Technical Conference (Usenix ATC '11)|
|Location||Portland, OR, USA|
|Abstract||Debugging operational networks can be a daunting task, due to their size, distributed state, and the presence of black box components such as commercial routers and switches, which are poorly instrumentable and only coarsely configurable. The debugging tool set available to administrators is limited, and provides only aggregated statistics (SNMP), sampled data (NetFlow/sFlow), or local measurements on single hosts (tcpdump). In this paper, we leverage split forwarding architectures such as OpenFlow to add record and replay debugging capabilities to networks – a powerful, yet currently lacking approach. We present the design of OFRewind, which enables scalable, multi-granularity, temporally consistent recording and coordinated replay in a network, with fine-grained, dynamic, centrally orchestrated control over both record and replay. Thus, OFRewind helps operators to reproduce software errors, identify data-path limitations, or locate configuration errors.|