email@example.com, Saltzer@mit.edu, firstname.lastname@example.org
This note comments on a current design controversy that can be framed partly in terms of end-to-end arguments. One form of active networking, a novel category of communication network architectures, comprises attaching programs to data packets, with the intent that those programs be executed at points within the network. The purpose is to better match the behavior of the network to the requirements of the application. The question being raised is whether or not this idea directly violates an end-to-end argument.
On the one hand, programmability appears to contradict the end-to-end principle that a function or service should be carried out within a network layer only if it is needed by all clients of that layer, and it can be completely implemented in that layer2. On the other hand, programmability may allow a network client to implement precisely the service it needs, an outcome that is consonant with end-to-end arguments. Contradiction and consonance aside, because programmability enables such a wide range of possibilities, applying end-to-end arguments in a general, yet definitive, way may be impossible. Instead, the specifics of each particular active networking idea would benefit from evaluation in light of the end-to-end principle.
At the same time, one can raise a concern that a general programming interface can lead to complex and unpredictable interactions among independently designed applications and independently acting users. Part of the context of an end-to-end argument is the idea that a lower layer of a system should support the widest possible variety of services and functions, so as to permit applications that cannot be anticipated. That is, minimize the lower-layer function, get out of the way, and let the higher layer do its thing. But this flexibility actually implies that end-to-end arguments have not one, but two complementary goals:
o Higher-level layers, more specific to an application, are free to (and thus expected to) organize lower-level network resources to achieve application-specific design goals efficiently. (application autonomy)
o Lower-level layers, which support many independent applications, should provide only resources of broad utility across applications, while providing to applications usable means for effective sharing of resources and resolution of resource conflicts. (network transparency)
While making lower layers more active or programmable is likely to enhance application autonomy, the risk is that programmable lower layers may reduce network transparency. The reason is that a key element of transparency is some ability to predict how the network will behave.
Since lower-level network resources are shared among many different users with different applications, the complexity of potential interactions among independent users rises with the complexity of the behaviors that the users or applications can request. For example, when the lower layer offers a simple store-and-forward packet transport service, interactions take the form of end-to-end delay that can be modeled by relatively straightforward queueing models. Adding priority mechanisms (to limit the impact of congestion) that are fixed at design time adds modest complexity to models that predict the behavior of the system. But relatively simple programming capabilities, such as allowing packets to change their priority dynamically within the network, may create behaviors that are intractable to model, in the same way that the simple rules of cellular automata such as Conway's Game of Life can lead to remarkably complex behavior.
To maintain the largest degree of network transparency, then, the end-to-end principle requires that the semantics of any active features be carefully constrained so that interactions among different users of a shared lower level can be predicted by a designer who is using the services and functions of that active layer. Lack of predictability thus becomes a cost for all users, including those that do not use the programmability features. Getting the semantics of active enhancements right is a major challenge, and wrong active enhancements are likely to be worse than none at all, since everyone helps pay the cost of something that is used by only a few but reduces transparency for everyone else.
Thus, even though active network ideas are not ruled out by end-to-end arguments, we have not seen practical, high-impact examples of a sufficiently simple, flexible, and transparent programming semantics suitable for use in lower levels of networks. Until such examples are developed in detail, the existence of active networks that meet the end-to-end criteria should perhaps be classified as theoretical.
An end-to-end argument is similar to the argument for RISC: it serves to remind us that building complex function into a network implicitly optimizes the network for one set of uses while substantially increasing the cost of a set of potentially valuable uses that may be unknown or unpredictable at design time. A case in point: had the original Internet design been optimized for telephony-style virtual circuits (as were its contemporaries SNA and TYMNET), it would never have enabled the experimentation that led to protocols that could support the World-Wide Web, or the flexible interconnect that has led to the flowering of a million independent Internet service providers (ISP's). Preserving low-cost options to innovate outside the network, while keeping the core network services and functions simple and cheap, has been shown to have very substantial value. In this way, an end-to-end argument does not oppose active networks, per se, but instead strongly suggests that enthusiasm for the benefits of optimizing current application needs by making the network more complex may be misplaced.
This idea of ``stupid operating systems'', ``stupid networks'', and ``stupid processors'' hasn't yet had its full run3. The telephone company still seems to think that all users want the illusion of a copper pair from the user's house to some ISP point of presence in another city. In its Internet access approach, the cable company places the real network one step closer, but neither architecture is really prepared for the household with three computers and two network access providers. Politicians want both the cable company and the ISP to filter packets for things children shouldn't see, and the FBI asks them to make copies of specific data streams to simplify wiretapping (these may be examples of non-cooperating end-points). Political arguments aside, even if one accepts these requirements, the corresponding implementation proposals are sometimes stunning in the way they fail to consider end-to-end arguments.
 D. L. Tennenhouse, J. M. Smith, W. D. Sincoskie, D. J. Wetherall, and G. J. Minden. A survey of active network research. IEEE Communications Magazine 35,1 (January 1997) pages 80-86.
 Elwyn R. Berlekamp, John H. Conway, Richard K. Guy. Winning ways for your mathematical plays. Volume 2: Games in particular. Academic Press, 1982. Pages 817-849.
 Michael D. Schroeder, David D. Clark, and Jerome H. Saltzer. The Multics kernel design project. Sixth ACM Symposium on Operating Systems Principles, in ACM Operating Systems Review 11, 5 (November 1977) pages 43-56.
 David P. Reed. Processor multiplexing in a layered operating system. S.M. and E.E. thesis, M. I. T. Department of EECS, 1976. Available as M. I. T. Project MAC Technical Report MAC-TR-164, July 1976.)
 D.D. Clark, K.T. Pogran, and D.P. Reed. An introduction to local area networks. Proceedings of the IEEE 66, 11 (November 1978) pages 1497-1516.
 George Radin. The 801 Minicomputer. Proceedings of the first ACM Symposium for Programming Languages and Operating Systems (1982), in Computer Architecture News 10, 2 (March, 1982) pages 39-47.
(1) Author's affiliations: David P. Reed, Techburst, 16 Schaffner Lane, Dover, MA 02030; Jerome H. Saltzer and David D. Clark, M. I. T. Laboratory for Computer Science, 545 Technology Square, Cambridge, MA 02139.
(2) There are some situations where applying an end-to-end argument is counterproductive. One category is costrelated. For example, many-to-many communications can be done by forwarding packets among multicast servers located outside the network, but it appears to be much more effective to add some support at the lowest layers of the network. Another category is where for some reason the endpoints are not in a position to cooperate. For example, version 5 of the Kerberos authentication system pulls inside the application programming interface the function of replay prevention, because experience with version 4 showed that few application programmers understood how to do that critical function correctly. The interesting thing is how few examples have turned up in 20 years of experience with systems like the Internet.
(3) David Isenberg of AT&T has recently re-invented end-to-end arguments in an entertaining paper called "The Rise of the Stupid Network" (http://www.computertelephony.com/ct/att.html) that criticizes the telephone industry's concept of "Intelligent Network" based on ideas very similar to those in the original end-to-end arguments paper.