Computing Logical Network Topologies

Abstract

It is increasingly the case that a network service will need to communicate over some type of logical or overlay network, built on top of a physical network. For example, in user-level multicast, there is a need to construct an overlay network to support one-to-many or many-to-many communication. With one-to-many communication, there is often a trade off between communication cost and end-to-end performance. At one extreme, one can obtain excellent performance by having direct pair-wise communication channels, hence foregoing the benefit of multicast. At the other extreme, a minimum spanning tree can be established among participants to minimize communication cost, but the end-to-end performance may be very poor. Past studies [Deering et al., 1996, Wei and Estrin, 1994] have clearly highlighted this trade off, and existing designs tend to favor lower communication cost because of scalablity. However, minimizing communication cost alone is not always enough to have a scalable system. If packet forwarding involves many nodes in a multicast network such as in a minimum spanning tree, it would be difficult to build a service because the service requires the deployment of many nodes with forwarding capability. Thus, for a scalable service, minimizing the number of service nodes should also be an objective.