Topics

Non-tech
CLID 71
TOPS 74
Backus 83
Bandw 96
Thesis
Patents

Dramatic Bandwidth Growth Implications

Eric Bierman, November 14, 1996

---

1. Introduction

Over the past few years there has been a strong interaction between the three driving factors in the bandwidth marketplace: the demands for bandwidth, the cost of bandwidth and the supply of bandwidth. It is our contention that these will continue to interplay until bandwidth becomes effectively free. This has serious implications for a number of industries, which generate both challenges and opportunities for them.

1.1 Demands for Bandwidth

The demands for bandwidth from the residence have been driven largely by the dramatic rise of interest in and in numbers of subscribers to the Internet. (Doubled from 6.2 million households to 14.7 in 1996 in the US alone). That is coupled with the increasing popularity of higher bandwidth applications (such as higher quality graphics and eventually video) on that same Internet.

In short, a growing segment of the residential population wants access and wants it at the highest bandwidth economically available. There is no limit to the consumer's desire to use the ever increasing bandwidth being made available by access competition, only a limit as to what he will pay for it.

A second demand for bandwidth at the residence is the desire for greater choice in broadcast video sources. The 500 channel universe seems small even before it is delivered. The desire to watch a specific channel in a specific non-local city is not yet satisfied.

This second need goes unfulfilled primarily because of the local cable monopolies today. It may well be first satisfied via the Internet.

1.2 Cost of Bandwidth

Between continents via undersea cable, across continents with multiple competing carriers, between cities, and within cities, whether by telephone companies, cable companies or alternate carriers, the medium of choice is fiber. Fiber speed (capacity) has moved from 500 Mb/s to 2.5 Gb/s to 10 Gb/s for basically the same cost.

The major cost component is typically the laying of the fiber. Once in the ground that fiber can usually be upgraded to the next higher rate by changing only the end terminal equipment, an incremental cost for a fourfold increase in bandwidth. Also, when plowing in the fiber, it is not individual strands of fiber, but typically cables of 24 pairs of fibers, yielding a route of 240 Gb/s capability in each direction.

On the distribution side, the last few miles to the individual residence, the low cost of fiber, is bringing that fiber closer and closer to the residence. This not only provides more bandwidth for lower cost on that fiber portion, but it also shortens the length of remaining copper to the residence, increasing its bandwidth capability and reducing its attendant costs (amplifiers and coding techniques) all while improving quality of signal.

1.3 Bandwidth Supply

With at least four major competing continental carriers in the US putting in place four individual fiber networks spanning all the major cities, all with 1/4 of a Terabit per second capability on each leg in each direction, we are looking at significant bandwidth supply capability. Additionally, virtually every electric power utility, railway and gas transmission company is getting into the act via their rights of way.

Competition is now creating the same growth of capacity under the oceans and in places such as Europe and Japan.

Within cities, telephone companies have moved entirely to fiber between central offices, because of its significantly lower cost compared to multiplexed signals over copper, providing yet another available pool of underutilized bandwidth for future exploitation. Cable companies are deploying fiber at a rapid pace to save maintenance costs of amplifiers and improve signal quality. Their other motive is, of course, to compete with the telephone companies in voice and data, and to stave off competition from them in video.

---

2. What is the Real Growth of Bandwidth?

We are all familiar with Moore's Law of fourfold improvements every three years for IC density, which leads to the same improvement rates for memory sizes, memory costs and processing power. George Gilder suggests that the next power curve will be for bandwidth, and it will run at ten times the Moore Law rate, however he does not give very convincing facts to back this up.

In the 60s we had 110 baud and 300 baud, in the 70s we moved to 1200 and 2400. In the 80s we jumped to 9600 baud, but by early 90s we were at 14.4, then 28.8, now 33.6 and soon 57.2 Kb/s. In 1995, with ISDN, we moved to 128 Kb/s and with ADSL next year, we can move to 7 Mb/s, or via cable modems, to 30 Mb/s. By the early 2000s we will have fiber to the home giving us 155 Mb/s or 622 Mb/s or more.

Looking at the recent push, we can say:

Once we get to fiber to the home, the rate of increase of bandwidth will follow a more modest curve, that of the fiber improvements we have seen over the past few years:

From this history, we can see a rate of 4 fold improvement every 10 years in the early days, increasing recently to well in excess of the 40 fold increase every 3 years predicted by Mr. Gilder. However, when we get to fiber-to-the-home, the Moore Law numbers will re-emerge. What may be more instructive is to look at the adoption rate for the various approaches to come up with an average rate of increase of bandwidth subscribed to in the residential market place. A summary of these four curves are shown in the following figure:

Using predictions of adoption rates of the various speed modem technologies as forecast by taking the most conservative of the August 1996 estimates of both IDC and Dataquest, we can get two curves, first, the growth of number of modems, by technology:

and, secondly, the total residential subscribed to bandwidth implied by these modem numbers and only a fraction of their eventual capability.

It is also instructive to look at the total subscribed to residential bandwidth in tabular form, in Terabits per second.

'96	0.9 Tb/s
'97	2.7 Tb/s
'98	7 Tb/s
'99	16 Tb/s
'00	33 Tb/s

The voice band modem figures, in spite of their large unit numbers, quickly get swamped by the ADSL and cable modems, even when rated at 2 Mb/s and 10 Mb/s (instead of 7 and 30) respectively, merely because of their significantly higher data rates. Hence the above numbers are highly sensitive to the unit forecast accuracy of those two technologies.

Given those numbers come true, it appears that over the years from 1996 to 2000, the bandwidth growth rate is fairly constant at a rate of about 15 fold every 3 years, much better than the Moore Law rate, but not 10 times it. Given the fiber to the home future, already starting in high density residential (apartment buildings) and new high income suburbs, the more than doubling every year could well go on for another 10 to 15 years.

---

3. Network Implications of Greater Bandwidth

Internet users tend to use anywhere from 5 to 100 hours per month, but a good figure seems to be about 30 hours, which equates to 1 hour per day or about 4% of the time. Of course there is strong peakedness of this traffic, which must be taken into account. Also, during this "on-line" time there is think time, read time, dream time, when there is no activity on the link; this could be of the order of 90% of the time, although there are bursts of non-stop activity when following links in a hurry. Perhaps a good overall figure would be 1% utilization of the residential subscribed to bandwidth.

1% of the 33 Terabit per second for the year 2000, is a more manageable 330 Gigabits per second. However, because of the constant demand pull, that more than doubling of subscribed to bandwidth will most assuredly be coupled with a more than doubling of used bandwidth. This drives the implications for the network.

The same 1% figure applied to the 1996 number of 0.9 Tb/s yields 9 Gb/s which is in good agreement with the 3 Gb/s derived from George Gilder's 250 Terabytes per month estimate on the backbone in July 1996.

Looking at audio and video stream usage, it appears to remain well below 200 Gigabits per second as late as 1999. In the year 2000 it ramps abruptly to 2 Terabits per second, but most of that will be either very local or multicast, which has a much lower impact on the network.

3.1 Backbone Load

While earlier it was pointed out that each of the North American continental carriers had "capability" for 240 Gb/s, some of them have only recently moved from 45 Mb/s to 155 Mb/s for Internet traffic on their backbones. Now MCI has just announced a goal of 160 Gb/s routes for its backbone by 1998, for merged voice, data and video over ATM. This, if dedicated to Internet (which it won't be initially), would allow MCI to carry half of all the traffic predicted above for the Internet by the year 2000, two years ahead of time.

Frontier, of Rochester NY, has just announced the start of implementing a 24 fiber pair network coast to coast, serving 100 cities, with 13,000 route miles, "surpassing all other carriers in bandwidth and reliability". The others will also have to turn their unimplemented "capability" into reality, and will do so as soon as they see MCI and Frontier picking up the advantage from this move.

3.2 The Router Load

In today's Internet, all the "nodes" of the network are routers, and with the world wide reach of the network, more and more Internet traffic goes through more and more routers to get from the source to the destination. As the traffic increases at greater than Moore's Law, the increase in speed of the routers cannot keep up with the increased traffic. This requires more routers, which can in turn lead to more routers in a source-destination path, increasing network latency.

One of the solutions to this problem is to increase the number of ports on routers, to generate more direct paths and result in fewer routers per path. This, in turn, results in uneconomical links over longer distances, which can be a different kind of problem. One solution to that problem has been to use virtual links (virtual circuits) over ATM fabric, where underutilization can be consumed by other virtual links.

A particularly ingenious solution is the IP switching technique over ATM fabric, promoted by Ipsilon. This solves two outstanding problems in communications at once, it creates a very high speed "router" easing the router load problems for some time, and it overcomes the awkward switched virtual circuit signaling, set-up and management problems in ATM networks.

With IP switching, Internet traffic goes through a router as normal, until a flow is detected. Then an ATM virtual circuit is set up to carry that flow. The technique keeps removing routers from a flow until there is as direct a connection as possible. This neatly solves the router load problem by reducing the number of routers in a flow and increasing the route diversity economically. As a significant "extra", IP switching works best on long holding-time "flows" such as the very audio and video streams which will be the main drivers of additional bandwidth use.

3.3 Separation of Streams and Control

As stream traffic increases, there will need to be a greater separation of streams and control. In the telephony world, that was accomplished in the networks by Common Channel Signaling. In the Internet, this appears to be more of a challenge. It could probably be done via "port numbers", allowing one port number to speak on behalf of another. In this fashion, a control TCP/IP connection could dictate what flows over other ports of the same IP number (not unlike what the ftp protocol does with separate "control" and "data" channels). The IP multicast upstream "joining" techniques are also illustrative of this notion. Ipsilon style IP switching already supports fully the IP multicast control protocols, so adding more separation of streams and control is probably a natural on this technology. The browser and server sides of this approach would also seem quite straightforward.

The approach leads to defining different qualities of service for different kinds of streams and, in turn, to making the Internet technique the single surviving mode of transport for all media.

The notion of voice and video flowing over ATM virtual channels, in a fashion not too different from how the cable and telephony worlds think of them, all while being controlled in the same fashion as web events take place today, seems to marry the best of both worlds in a single paradigm.

What this means to the Internet world is that the increasingly popular streams (Real Audio, Real Video) will gradually move off TCP/IP, then off IP, eventually onto raw ATM virtual circuits, but controlled by a related TCP/IP connection.

What this means for the telephone world is the elimination of signaling protocols, as they get replaced by browser style control connections, personalized directories, and merging with the non-telephony services.

For the cable world, browser style selection of channels means that the channel selection can take place within the intelligent TV set, within the set-top box, or within the network, so that the selection range is no longer limited to the coax cable limitations (40, 70 or 110 channels). This makes the single channel ADSL pipe from the telephone company as wide as the coax or DTH satellite when it comes to selection.

3.4 Control Distribution

While the telephony switching world represents an amazing example of distributed control, with excellent redundancy and hence reliability, there are many legacies of its evolution which keep getting in the way. The fixed numbering plan which now has to be freed up with local number portability solutions, the voluminous data filling required to update routing tables and translations in each of the many switches, the variety of ever more complex signaling protocols used to forward calls and query databases and the elaborate algorithms required to reroute around network outages are but a few examples.

The Internet has the advantage of starting more recently with a different approach. Routers more dynamically become aware of their surroundings and configure their routing tables algorithmically on their own, and reconfigure them on the fly to handle network outages using those same algorithms. Each router needs only to be datafilled with information on its immediate neighbours, and from then on, operate independently, while part of a vast network.

One of the truly elegant aspects of Ipsilon's IP switching approach is that it maintains that router style independence, in effect extending it to ATM switches. A network, consisting of independent ATM switches, each individually controlled by its Ipsilon switch controller (router), becomes a high powered, low maintenance network. It becomes an almost indestructible network that is easy to grow or change and does not need a single owner or manager. It reflects the "organized chaos" of today's Internet. The control is fully distributed.

3.5 Server Activity

As more users come on board, and as video and virtual reality files become more popular, and particularly as subscriber bandwidth increases, the server response must increase. The current approach of riding the Moore Law curve to provide faster and faster servers is not going to keep up with the faster bandwidth growth.

In addition to symmetrical multi-processing server enhancement, Cisco and others are providing mechanisms for distributing requests on a specific IP address to a multitude of servers, which need not be co-located, to handle the load. This works particularly well, because most web hits are "stateless" or "independent", hence it does not matter which server is hit with a request or the subsequent requests from the same client. The multiple copies need only be kept very loosely identical.

This allows the two-fold solution to the faster growth rate of using both faster servers and more of them. It even adds greater "availability".

Another significant helper in reducing server load is the use of IP multicasting for any live source, such as real-time sports events, political events, NASA launchings, etc. A server can get away with a single stream out, letting routers along the way build the trees to reach large numbers of recipients with the same content at the same time (loosely). This technique also has the potential to distribute cable TV type content over the soon to be fatter Internet.

Another approach will be the separation of "control" and "content" within the server structure, where the requests go to the "control" engine and the responses are generated by the "content" engines, which could be multiple engines within the "system". The Tandem "ServerNet" approach to video servers could be made to operate in this fashion for normal web requests as well, particularly when the requests are for voluminous chunks of data, such as video clips or large software downloads. The increasing use of video and audio "streaming" rather than bulk file downloads will also push the network to separate control and content approaches.

3.6 Traditional Telephone Switching

Now with data traffic finally exceeding voice traffic in the network, the traditional telephone switch, handling 64 Kb/s streams, will gradually become irrelevant. Already Nortel Internet "Thruway" boxes are being put in front of the line cards to extract Internet traffic. ADSL also intercepts the copper pair to the subscriber to put on (pull off) the 7 Mb/s down, and 500 Kb/s up data streams. These traditional switches cannot interface directly the 2.5 Gb/s streams on the trunk side either. In fact a typical voice switch only switches a grand total of less than 5 Gb/s, so the entire switch could be replaced by a "remote" off an ATM switch somewhere via two 2.5 Gb/s fiber pairs.

The whole essence of "call processing" for the subscriber can disappear into browser plug-ins. Features can be defined by the user or by a shrink-wrap software developer. Network services, such as 800 numbers and call centers can be provided by a combination of DNS servers and other servers. Conference calls can become IP multicasts. The distinctions between a voice number, a fax number and an e-mail address can disappear, and all the "calls" or "conferences" can be multi-media. It has an irresistible attraction.

However, the switching of audio or video streams must be done within the human equivalent of "instantaneous" of about 200 milliseconds. This requires the IP multicasting "join" functions to operate at very high speed, very similar to switch "call processing", within the routers or IP switches. It is here that traditional call processing heritage and core skills can be applied.

3.7 PC Bottlenecks

As the use of bandwidth at the home increases faster than the processing power and memory speed, the processor and the buses and the memory will become bottlenecks. Already the faster increase in speed of the processors compared with the speed of low cost memories has resulted in two levels of caching, with its attendant problems. What starts to be necessary is for streams to go directly to where they will end up, bypassing memory to go directly to screens or co-processors.

This again leads to a separation of streams and control, where the control sets up the appropriate multiple paths for the various streams. The main processor becomes an overall controller for directing the streams to, say, MPEG-1 or MPEG-2 co-processors or whatever is required. If the buses can keep up, a single bus will keep the costs down, but if not, separate buses will need to be created cost effectively.

Going the "network computer" route allows removing some traditional PC costs to allow adding of parallel stream costs. Remember, with high speed, low cost network connections, the need for local disks will disappear, as access speeds to network based disks will be comparable to local disk access speeds. Access to the network will become more integral to the computer architecture.

---

4. What Others are Doing

Digital Equipment Corp

DEC has for years been involved in router development, as part of its long term development of DecNet. It has, as a result, people skilled in the lore of router technology, is both knowledgeable and opinionated about the many variants of IP routing philosophies and has strong connections with other router vendors.

Dec has for a long time also targetted the telephony marketplace for its products and has become knowledgeable there too. It has a position in the Intelligent Networking (IN) market, and this knowledge has spilled back into their data areas. Dec is an active driver of Virtual Private Network techniques in the data area via IP tunneling.

Dec is heavily involved in ATM, with a 10 Gb/s ATM switch (with optional credit based flow control), GIGAswitch, and SAR chips for other people's ATM cards. It has become an Ipsilon "partner" and is well positioned to build IP switches, given its expertise in both routing and ATM switching.

Dec is in both the PC and Alpha workstation businesses and has significant expertise in building Vax clusters as powerful servers. It is very active in the Internet business, through its Altavista search engines based on the high powered Alpha server clusters.

IBM

IBM has for some time now championed ATM, and promised that they would be delivering ATM switches in a big way. So far little has emerged, but their heavy push for 25 Mb/s ATM to the desktop has given them a big voice at the ATM Forum, and now, after they have succeeded with 25 Mb/s, they are giving lip service to 155 Mb/s. They are spending a lot of money on ATM research which could lead to big switches in the future.

IBM offers a Global Network for voice and data communications to businesses. At the same time they are very strong in voice recognitions systems and voice dictation systems. Their current efforts at ITU for standardizing voice over IP for the internet may point to a strategy for the future.

In the Internet business, IBM is somewhat weak, with their second class WebExplorer, and some other tools, but they are there. Their PC business seems to be reviving well, as is their server business with a variety of PowerPC/RS-6000 based platforms.

Fore Systems

We are aware of Fore's position in the ATM marketplace, but it should be of interest that Fore is an active Ipsilon partner and, in fact, is working with Sprint to help them develop a national IP service over ATM with Fore switches and Ipsilons routers. This latter arrangement specifically excludes us even though we have a partnership with Fore and are a major supplier to Sprint. Does this mean Sprint feels we are their favorite supplier of yesterday's products?

Lucent Technologies

The company we feel we know the best and we have competed with the longest, has made interesting inroads into areas related to the bandwidth explosion. Their current release of Inferno builds credibility in environments for set-top boxes and network computers as well as riding the byte-code wave generated by Java.

At the same time Lucent has aggressively entered the Internet voice and fax business with software encoders and decoders for voice, fax, CD quality music, video and even echo cancellation through its elemedia arm. Their microelectronics arm is producing chips for the latest popular modems and has a single chip MPEG-2 decoder for set-tops, network computers and PCs. On the protocol side, they have made significant contributions toward reliable IP multicasting and we know they are actively looking at multiple separate stream mechanisms.

They are not at all hesitant to encourage bypassing the traditional carriers via voice, fax, music and video on the Internet, either via their PBX line of business or directly to the ISPs. We would do well to look beyond what they are doing with their legacy products.

Newbridge

Newbridge has not embraced Ipsilon IP switching that we are aware of, but are instead still pursuing their VIVID approach to extending LANs to WANs and achieving the scalability necessary for "greenfield" applications. This does, of course, point to their good knowledge of routing and how IP networks behave. However this approach will not work to help the existing Internet towards growth and scalability with easy interoperability. It will be interesting to see how they fit into the evolution of the Internet while maintaining the VIVID approach, or if they will modify their approach to enter.

Cisco

Cisco has recognized that their routers will not be able to keep up with the rapid growth of Internet traffic, particularly with audio and video streams, and are using their weight at IETF to push "tag switching", a half step to the equivalent of IP switching and are pushing for the "rsvp" protocol to provide some quality of service.

These are likely the typical marketing ploys of the dominant supplier to keep people away from IP switching until they have the equivalent ready for product. If Ipsilon keeps up the pressure, Cisco's dominance will be at risk.

---

5. Possible Scenario

Voice over IP on 28.8 Kb/s is quite good. Audio over IP on 28.8 is better than short-wave radio. Stereo over IP on ISDN is quite good, and CD quality on ADSL. MPEG-1 video over IP on ADSL is quite good, although not "broadcast quality". MPEG-2 video over IP on a 30 Mb/s cable modem will be broadcast quality and will compete directly with MPEG-2 broadcast video over the same cable.

It seems quite likely that the unifying concept of everything over IP will win, primarily driven by the Internet momentum and by the capabilities of the ever increasing bandwidth. The user interface for all this will be web-like, and evolve in interesting ways over time. Think about "clickable" areas on a live video, audio programs always having a clickable visual component.

IP streams will flow over ATM in the network controlled by IP switches. Content and control will gradually separate into separate but related IP streams, going to (coming from) different components of compound servers. The server components will specialize to perform their role best.

Over time, the content IP streams may move off IP onto more efficient flatter protocols, but this will definitely be a later development, as it affects many components.

In the home, many appliances will plug into a home LAN via a single common cheap interface. The TV sets, the radios, and the telephones will be first, as the content "browsers". Subsequently, many lower bandwidth devices can be added as mini-servers to allow their monitoring and control by the first set of appliances.

IP can be the single unifying medium in the coming world of low cost bandwidth, and if it can, it probably will be.

---

6. What We Must Do

If a vision makes sense after some intense analysis and discussion, it should be adopted and used as a guide for the many technical and investment decisions made every day within the company.

There are three major areas where we can start to make this vision happen, and they can be loosely related to our lines of business. They are "Consumer Products", "Infrastructure Products" and "Mindshare".

6.1 Consumer Products

(a) We make a variety of "telephones" and they have a variety of interfaces, some public and some proprietary. In our digital wireless phones we already packetize the voice and ship it via an air interface. What if we were to wrap these voice packets into IP frames. What is the marginal cost?

(b) We could build wireline phones in the same way, packetize the voice, wrap it in IP, leave out the air interface. This would cost more than the basic Jazz telephone set, but it would be much cheaper than a cellular phone and could have interesting features that sell it. Think also of the line card cost savings in the central office.

(c) When we build an Orbitor or a Vista, we have a voice stream and a data stream and worry about simultaneous or alternate voice and data. But, if we wrapped the voice in IP and called it data, we would solve that problem.

(d) ISDN in the home was an embarrassment, because it would not work with ordinary phones without an expensive adapter. Now ISDN is going in, connected to a mini-router for Internet access from your PC. If we extend that, IP stream phones could be attached as extensions at will. A home LAN, initially twisted pair ethernet on ISDN and on ADSL and on cable modems, is a potential product.

(e) The Orbitor and the Vista wired phones, Java enabled, can easily become a network computer. When large, 40"+, flat panel displays, get cheap enough, the network computer could turn into the home entertainment center, possibly getting us heavily into the consumer marketplace.

6.2 Infrastructure Products

(a) As we place AccessNodes in front of traditional switches to pull off Internet traffic, we could instead pull off all traffic, routing it to appropriate servers as needed. In fact, once the subscriber streams are all IP wrapped, both the AccessNodes and the Rapport systems become simple small IP switches and all we keep of the access side are the appropriate multiplexed physical interfaces.

(b) In the central part of the network the switch becomes the IP switch, the router controlled ATM switch. Hence all of our ATM switches need to be controllable via the IP switching mechanism. This will need to be enhanced from the Ipsilon defined GSMP to include quality of service mapping and to allow initially the joint existence of IP switching and traditional ATM switching with the appropriate bandwidth sharing through quality of service parameters. Eventually this will evolve to pure IP switching, when the premise of this paper fully takes hold.

(c) Similar IP over ATM mechanisms can be used to create high powered server islands, many servers with a single logical IP address or with multiple IP addresses for different services within a large single administrative island. Some have called this an IBX or Internet branch exchange. This certainly helps with the scaling issues of the Internet.

(d) Voice gateways need to be created to allow the connection of IP wrapped voice to existing telephone networks. Servers will also be required, for a while, to act as voice converters between different formats of IP wrapped voice, although this need should die out quickly as the consumer devices change to support most of the popular formats in use, or a single format wins out over time.

(e) Existing servers such as voice messaging servers will need to convert to accepting and delivering IP wrapped voice, to give them universal accessibility. This may well be a lead product because of its immediate utility in giving IP voice access to network answering machines for ordinary voice terminals.

(f) The small access IP switches when used to switch video streams for a large number of subscribers attached to it, will need to keep up with rapid channel switching. This will probably require it to have pulled into it a server function for channel selection with a performance far greater than achieved by most servers. This will open an opportunity for us to apply some of our real-time call processing skills to this server operation. Additionally, our expertise in high reliability operation will play a role in making us masters of this opportunity.

(g) As multicast content streams become a major component of Internet traffic, that separation of control from content streams that IP switching allows, will allow both IP switches and content servers to adopt a different structure where the content component can be optimized to deliver or pass long packet content, while the control component can be optimized to do short packet, real-time call processing style control over the streams and sources. This also calls for an addition to the Ipsilon method of flow detection to allow streams to be defined as flows from the outset, saving some load from the router/controllers.

6.3 Mindshare

To lead our way into these product opportunities, we need to become as dominant in the Internet Engineering Task Force and the ATM Forum as we have been in the ITU for the past 30 years.

(a) In the ATM Forum, we need to spend less time on the SVC signaling topics and voice over ATM topics and more time on the IP switching topics, aggressively proposing modifications for quality of service, for voice over IP over ATM, for flow definition in IP switching, re-argue credit flow control, etc.

(b) In the IETF, we need to become more active, through our Passport product, on topics such as voice over IP, video over IP, IP multicasting, quality of service, IP switching over ATM and separation of streams from control, all while downplaying the myriad of alternate methods of IP carriage over ATM.

(c) We need to embrace publicly the IP switching mechanism, both with Ipsilon and as a more general mechanism. We need also to roll out a few small products as identified above as quickly as possible, to be followed later with more significant ones.

(d) We need to engage University researchers in exploring some of our ideas and listening to some of their ideas in these areas.

(e) A road map can be developed quickly internally to achieve consensus, which can then be externalized with selected customers and finally publicly.

---

7. Conclusions

We seem to be at one of those discontinuities in our business, where dramatic shifts will occur in how our infrastructure customers deliver their services. We have faced these before and have aggressively driven our customers through the change. We need to prepare to do this again!

---

© Eric Bierman, 1996.

November, 1996.