PLAN and PLANet: An Active Networking Testbed

			Scott Nettles
			Dept. of Computer Science
			University of Arizona

The goal of active networking (AN) is to greatly increase the network's
flexibility and to facilitate the rapid evolution of the network
infrastructure.  Our attack on this problem is to make the network
programmable. Ideally, such an approach combines research into both networking
and programming languages. To that end, we have designed and implemented a
special-purpose programming language, PLAN (Packet Language for Active
Networks), and used it to build an internetwork, PLANet. Together they form a
testbed that allows us to explore the AN design space and to put some upper
bounds on the expected performance of such systems.

There are two basic ways to provide programmability: one, by allowing
extensions to be dynamically downloaded into the network infrastructure, and
two, more radically, by making packets into programs and allowing these
programs to be executed as part of network operation.  We are exploring both
approaches. PLAN serves as a limited, but safe and secure, packet language,
while downloadable switchlets written in the Caml language provide for fully
functional, but possibly heavyweight, infrastructure extensions. PLANet is an
active internetwork built using the PLAN and switchlets. In PLANet, all packets
are PLAN programs and the basic infrastructure is built from switchlets.  New
network functionality is created by downloading switchlets into routers and
then using PLAN programs for distributed communication.

My talk will cover the design and implementation of PLAN and PLANet, as well as
present the results of several experimental studies. The first study looks at
performance on traditional networking metrics and shows, for example, that
PLANet can route packets at more than 47 Mbps on a 100 Mbps Ethernet. The other
study is designed to explore how flexibility can help cope with a standard
networking problem: congestion. I will present several approaches to coping
with this problem, the most radical of which uses PLAN packets to search the
network for uncongested routes and then routes around the congestion.

At the end of my talk, I will touch briefly on my work in high-performance
garbage collection and persistence/transaction systems for Standard ML.