Edging toward a distributed future

Transfer of power to the network's edge hinges on always-on connections and time-to-market guarantees

Talking futures is something we all enjoy taking a shot at, especially with the wealth of potential material currently available in the telecom industry. Given the amount of growth that communications networks are experiencing - across many terabits of traffic and the sheer diversity of applications - theories abound as to the future of the network core and its role.

Numerous well-documented drivers account for the changes, including the emergence of the Internet and intranets, multimedia applications, increases in desktop computing power and performance and corporate globalization. The net result is that baseline requirements are shifting - particularly in regards to bandwidth management, quality of service and the introduction of new scalable architectures for the future. Today the edge is the main focus of attention.

It is almost an axiom of current thinking that the network is evolving to a "stupid" core, to repeat David Isenberg, author of the 1998 essay "Rise of the Stupid Network: Why the intelligent network was once a good idea, but isn't anymore. One telephone company nerd's odd perspective on the changing value proposition."

In this model, intelligence resides at the edge of the network, leading to a more pleasing combination of greater innovation and chaos. In turn, this would presumably involve a transfer of power - from the big monopolistic phone companies to the masses.

But all this talk of the edge misses one fundamental point: What does that edge really look like? Is it truly a knife-like boundary between the phone company and the customer, as with customer demarcation points of old?

In fact, the "edge" really is distributed by nature and the processing associated with it will be shared between the customer premises equipment (CPE) and the access node. At first sight, this argument may appear to violate the "stupid" trend, but that is not necessarily the case.

Toward always-on service

There is a clear drive toward a distributed edge rather than a clean separation between the network and the CPE, and this is coming from many sources. The first trend to notice is the migration of the customer connection itself.

Considering the logical characteristics first, the phone connection has been an on-demand service for many years. The user picks up the handset to request a circuit from the phone company and then, by dialing a phone number, causes a connection to be set to the desired called party. The connection does not exist until the off-hook and dialing requests (or demands) are made and processed, making this an on-demand service.

As a direct result of the migration to packet networks and the Internet, services are becoming always-on instead of on-demand (Figure 1). The connection between the CPE and the network gateway can be permanently set for DSL or cable modem applications, and this is typically how it is configured.

In addition, considerable evidence exists of rabid voiceband modem users dialing their ISPs once and staying connected for days. There is still an on-demand element to this case, but it now applies to sessions on top of an already-established connection. You can have multiple simultaneous sessions (e-mail and several Web pages for example), but they come and go at the behest of the user. Figure 2 shows examples of some always-on network applications.

All this has a fundamental bearing on the evolution of the network, as the physical layers of the last-mile connection move from switched-access feeds to semi-permanent feeds. As a result, the access node to the network then becomes semi-permanently attached to a particular CPE or user. This, in turn, means that the millions of instructions per second (MIPS) and storage in the access node become a resource available to the CPE for several network and third party-based services.

This transition from switched access to semi-permanent association is a key transformation in network architecture, fundamentally altering the way in which network resources are managed. They could, in principle, be made available even for switched access, but issues of initialization and synchronization of databases generally make it prohibitive.

The fact that a section of the access node is semi-permanently attached to a particular CPE now means that the synchronization and initialization steps need be done only on connection establishment - in other words, on relatively rare occasions and not with the establishment of each session.

Other issues

The next trend to take into account in the drive to a distributed edge network is the issue of user information. For years, everyone's user interface was identical (at least within a geographic region). The Internet has begun to change that. A primitive example is the use of cookies to tailor a user's interaction with a particular Web site. In the future, as the access node becomes semi-permanently attached to a particular CPE, user information, including preferences, can be stored in the access node and the CPE.

Users will, of course, be able to select that portion of the information to be stored in the CPE instead of an access node. The user also will be able to migrate this data to new carriers.

There are two major advantages of this distributed storage. First, the user is protected from losing all the data if the CPE is damaged, lost or stolen. The network can restore the data to the new CPE, and there would be minimal disruption of service. Second, the services offered to an end user now can be more complex. Networks no longer need to be constrained by the least-capable CPE that a user might have, as the access node's capabilities can buffer this CPE variation.

Service providers also must consider the limitations at the edge itself. Today, there clearly is a bandwidth bottleneck in the access network. We tend to assume that this bottleneck disappears with DSL and cable access. This is a mistaken view - one merely has to look at the core network to see why.

Large companies such as Akamai and Inktomi base a considerable amount of their business offerings on buffering the Web against the huge bandwidth demands of the collective user community. If the core network cannot keep up, what makes anyone think the access link can? Actually, applications will rise to fill any void in bandwidth usage that may temporarily exist. The access node, therefore, will retain some of the role it has today in managing that bandwidth limitation.

The access node likely will take on another role, too, as the connections become semi-permanently attached, buffering MIPS limitations and variations in the CPE.

One very real barrier to extensive third-party application deployment is that there is a large variation of CPEs to design for. One either has to pick the most popular model, design for that and be content with the market share that results or deliberately dumb down the application to ensure that it can run on many CPEs.

However, there is no real need for this constraint to exist. If third parties (and network service providers) use the MIPS in the access node itself to buffer for the CPE's limitations, the problem can be overcome. Thanks to the semi-permanent attachment of the access node to a particular CPE, it is possible for the CPE to be seen by the application developer as a distributed machine with MIPS and storage, available at the customer premises and in the access node.

The applications developer, therefore, no longer needs to be concerned with the lesser amount of MIPS and storage available in older CPE. Network designs can use the MIPS and storage capabilities of the access node instead, thereby gaining considerable advantages in performance and scope.

Many other factors further reinforce the argument about the move to a distributed edge, such as scalability concerns and customer-driven service creation, but the impact of the always-on connection and resulting semipermanent attachment of the access node to a particular CPE represent the key factors. Sun Microsystems Chairman and CEO Scott McNealy's lonely cry for many years has been that the network is the computer - and he is about to be proved mostly right. The future will have the processing power of the access node available as a natural adjunct to the user's CPE. As such, the access node will be a computational resource available exclusively for the users' tasks.

This does not mean that the CPE will disappear entirely. Issues of privacy and speed also will continue to fuel an extensive CPE market. The access node will be used across the spectrum of CPE (dumb to high-powered), and users with limited CPE will be able to access the same services as their high-powered CPE colleagues, at the price of some speed and usage cost limitations.

Stupid is as stupid does

So what happens to the "stupid network" in all of this? One could argue that if the access node is owned by the network provider and becomes an adjunct to the CPE for new service creation, intelligence is just added back into the network. On the surface, it would seem that this distributed edge migration is a violation of the principles of the stupid network. However, a closer examination reveals it does not violate those principles.

There are several reasons for this. First, the programs that might run in the access node are under the control of the user. If those application programming interfaces (APIs) are not opened by the carrier to third-party service developers, the users will migrate to a more open carrier or simply bypass the access node MIPS in favor of a semi-permanent session connection to a third-party developer. While the latter is not as efficient as an implementation in the access node itself, it is a pathway to bypass the service provider that overplays that advantage.

Second, it is fair to say that the natural outcome of all this is to take advantage of the MIPS and storage capabilities in the access node to accelerate service innovation and deployment. However, it is not necessarily the only way to proceed. Services could be deployed by simply leaving the access nodes as they are now: dumb. It might not represent the smartest market move, but it is an option. By opening the access node to use, service providers can obtain efficiency in development and deployment of services and in execution of an individual service.

The market change drivers are clear: Customers want more services more quickly, efficiently and cost-effectively. For providers, the goal therefore in the new world must be to improve the time to market of new services. To achieve this, they have to expand the pool of innovative thinkers and developers. It stands to reason that a service provider that tries to constrain innovation by limiting availability of the access node MIPS (by keeping the API closed, for example) will lose out as service innovation grows more quickly on the competitive providers' networks and entices the monopolistically-minded providers' customers away.

We can argue that the stupid network trend is occurring primarily because of a drive to reduce time to market - the all-important objective for providers today. To gauge accurately whether any of the new trends discussed here violate the principle of the stupid network, we need to examine what the change does to the time to market for new services rather than argue about network architecture. If the change further reduces time to market, it is consistent with the stupid network principles; if not, it runs counter to it and ultimately will lose.

In the case of access node MIPS being made available to a CPE for use in running a service, the time to market for new services is definitely reduced. This statement is made on the basis that it is much easier to design for one API (the access node plus myriad CPEs) than for many different API's (one per CPE type and configuration). It is clear, therefore, that the concept of the distributed edge is fully consistent with the drive to the stupid network.