An important question about active networking concerns the relationship between the active processing paradigm and the more traditional end-to-end processing paradigm. The well-known end-to-end argument (Saltzer, Reed and Clark) is an architectural principle that guides the placement of functionality in a distributed system. We believe that active networking is a natural consequence of the end-to-end argument, because certain functions can be most effectively implemented with information that is only available inside the network. This position, and technical arguments to support it, are described in the paper "Active Networking and the End-to-End Argument" to appear in ICNP'97 at the end of October.
In collaboration with AT&T Labs, Research, we (Bhattacharjee) developed the Control-on-demand (C-o-D) architecture, which allows network users to introduce per-flow application-specific control policy into the network. User policies are dynamically downloaded into the network on demand (e.g., at flow initiation). Control-on-demand is a language independent approach towards active networking. Rather than specifying a specific language that must be used to program the network, we define primitives and abstractions that can be used to specify interesting policies. In order to improve the scalability of the system, we have developed a "frame peeking" mechanism to reduce the data volume that crosses the forwarding engine and the controller boundary. Further, we apply the on-demand control on a best effort basis to reduce on-line requirements on the control processor(s). A paper on this work was submitted to Infocom'98.
An in-kernel Linux implementation of the C-o-D architecture was developed at AT&T Labs in which Control on demand is provided for IPv6 flows. Several flow specific controllers for media-specific congestion control and multicast were been developed to test the architecture, and the controller-network interface. The implementation has been installed at Georgia Tech for continued work on the project.
In the area of inter-activity, we installed the ANTS active networking platform from MIT on a system at Georgia Tech. We have also begun plans for a small group meeting involving MIT, Georgia Tech, and a few other groups with common interests.
We are planning to submit a paper to the IEEE Network Magazine Special Issue on Active Networking. This paper will discuss the requirements of an active networking architecture, and describe our methods for meeting these requirements.
We are planning the next version of the active networking simulator AN-Sim. The new version will increase the flexibility of the simulator to allow modules to be easily added that describe basic functions such as the node processing model. These extensions will allow modeling of heterogeneous networks that contain both active and non-active nodes. The new version will also allow other research groups to simulate their AN processing models.
We will explore the possible use of a Washington University Gigabit Network Kit as a platform for some active network development. Our group will have access to two kits, when the distribution is ready (approximately 1Q98). During the next quarter, we will assess the capabilities of the kits and determine the types of active processing which can reasonably be included.
Building on Bhattacharjee's work on control-on-demand, we plan to continue developing architecture and mechanisms that i) support formal reasoning about active nodes' behavior, individually and collectively; ii) provide a simple set of building blocks for resource management; iii) support fast-path implementation techniques.