Service introduction in an active network

ion). As with Softnet, the messenger system is similar to the capsule approach. It demonstrates that communications protocols can be successfully structured as migrating computations by implementing a sliding window protocol. Further, it shows that with care the resulting programs can be compact enough to be transferred as datagrams. For example, a network discovery program is given in a terse 200 bytes. Unlike the system proposed here, performance was not a goal of the prototype system, and so the impact of messenger-style protocol implementations on overall application performance is not known. The overhead of code distribution is also unknown, though provision is made to refer to other programs installed at hosts rather than to always carry complete programs. 6.3.3 ACTIVE IP The active IP system was a proof-of-concept active network intended mainly for network probing and discovery tasks. It allowed network users to tag their IP packets with fragments of Tcl code. This code was carried by using the existing IP options mechanism. In this manner, packet processing at enabled nodes was customized in terms of a set of available primitives. The system was implemented by extending and embedding a Tcl interpreter in the IP layer of the Linux operating system kernel. Further details can be found in [Wetherall and Tennenhouse, 1996]. As a predecessor to ants, active IP demonstrated the feasibility of an active network for a restricted case. It showed that interesting and useful programs (such as the existing options tasks and network discovery programs) could be composed from a relatively small set of node primitives. It showed that these programs could be compactly expressed, such that it is reasonable to transfer them along with packet data. 6.3.4 Switchware and PLAN At the University of Pennsylvania and Bellcore, the Switchware project [Alexander et al., 1998] is developing a programmable switch approach that allows digitally signed typechecked modules to be loaded into a network node. Out-of-band program loading is used to support value-added services, such as network striping. This approach builds on the Advanced Intelligent Network (AIN) concept of the telecommunications industry, in which the implementation of value added services is separated from that of switching by moving the service control functions to adjunct processors. In Switchware, formal methods are applied to assure the security of the network; one goal is the identi cation of security properties of the underlying infrastructure for which theorems can be proved. A prototype \active bridge" demonstrates how network elements can be upgraded on-they [Alexander et al., 1997]. A complementary part of the e ort is the development of PLAN [Hicks et al., 1998], a programming language for active networks. PLAN is intended to be compact, so that small forwarding programs can be carried directly in each packet. These programs can refer to node resident code for privileged operations or common-case processing that is too large to carry directly. PLAN is also designed to facilitate the safe operation of the network. 139 Because the language is restricted, certain program forms, such as those that loop forever (either locally or across nodes) cannot be expressed. 6.3.5 Netscript The Netscript project [Yemini and da Silva, 1996] at Columbia University is focusing on network management. It is designed to support new routing, packet analysis, signaling and management tasks. Netscript consists of a data ow style programming language for scripting agents that process packet streams, along with a Virtual Network Engine, or execution environment within which agents are run at nodes. An overall program may consist of many agents that are distributed across nodes. Nodes themselves are connected via the underlying physical network by an overlay of Virtual Links. The system allows new agents to be dynamically deployed and con gured using a delegation model. The goal is to enable the programming of remote nodes, including intermediate systems, as easily and quickly as end-systems. Compared with ants, the Netscript approach is aimed at network managers rather than users, and hence the establishment of more heavyweight network processing con gurations. 6.3.6 SmartPackets and Sprocket At BBN, the SmartPackets project5 has also applied active network techniques to network management tasks, but for signaling and control rather than to introduce new functionality. SmartPackets carry short programs written in a specially-designed RISC-like language, Sprocket, that is compact and targeted at network management tasks. For example, MIB names are represented compactly. These programs are then authenticated and interpreted at each router they encounter to carry out SNMP-like management tasks. The basis for organizing the system in this way is the same as that of programmable RPC: that performance increases stem from the collapse of multiple round-trip queries into a single program execution at each node. 6.3.7 Active Services The active (distributed) services approach is based on the availability of programmable end-systems within the network infrastructure [Govindan et al., 1998]. These end-systems can then be organized into service-speci c overlays. Amir provides an example of an active service framework, AS1, targeted at media processing tasks such as transcoding [Amir et al., 1998]. One of the main motivations of active services is to provide active network-like functionality without requiring modi cations to routers. As part of the design of ants, I have argued that the distinction between end-systems and routers is not substantial in this context. Instead, the signi cant design constraints (common to both ants and active services) are that programmable nodes exist within the network infrastructure (and so are in a strategic 5Smart Packets for Active Networks, B. Schwartz et al., January 1998. From http://www.nettech.bbn.com/smtpkts/smart.ps.gz. 140 location to support new services) but that not all nodes need be programmable (and so there is no performance penalty imposed on all routers). A more substantial issue raised by active services is how to accommodate computationallyintensive processing, such as video transcoding and Web caching, within the network infrastructure. A reasonable view is that tasks fall into at least two domains: lightweight processing akin to ants forwarding routines; and heavyweight processing akin to AS1 data manipulations. Support for both domains is required and can work in synergy with each other. This view is supported by the exploration of services such as Web caching and the evaluation of the ants architecture. ants services act as the network \glue" that knits together processing that occurs at programmable end-systems into a useful overall pattern, and programmable end-systems extend the range of tasks that can be accomplished with ants services. 141