Large-scale Virtualization in the Emulab Network Testbed

Mike Hibler, Robert Ricci, Leigh Stoller, Jonathon Duerig, Shashi Guruprasad, Tim Stack, Kirk Webb, Jay Lepreau

University of Utah, School of Computing
emulab.net www.flux.utah.edu

Abstract

Network emulation is valuable largely because of its ability to study applications running on real hosts and “somewhat real” networks. However, conservatively allocating a physical host or network link for each corresponding virtual entity is costly and limits scale. We present a system that can faithfully emulate, on low-end PCs, virtual topologies over an order of magnitude larger than the physical hardware, when running typical classes of distributed applications that have modest resource requirements. This version of Emulab virtualizes hosts, routers, and networks, while retaining near-total application transparency, good performance fidelity, responsiveness suitable for interactive use, high system throughput, and efficient use of resources. Our key design techniques are to use the minimum degree of virtualization that provides transparency to applications, to exploit the hierarchy found in real computer networks, to perform optimistic automated resource allocation, and to use feedback to adaptively allocate resources. The entire system is highly automated, making it easy to use even when scaling to more than a thousand virtual nodes. This paper identifies the many problems posed in building a practical system, and describes the system's motivation, design, and preliminary evaluation.

In Proceedings of the 2008 USENIX Annual Technical Conference, pages 113–128, Boston, MA, June 2008.

• PDF