A funny thing happened on the way from the router.

We are all such good students of Moore's Law, the notion that processor speeds will double every year and a half or so, that in any digital arena, we have come to treat it as our 'c', our measure of maximum speed. Moore's Law, the most famously accurate prediction in the history of computer science, is treated as a kind of inviolable upper limit: the implicit idea is that since nothing can grow faster than chip speed, and chip speed is doubling evey 18 months, that necessarily sets the pace for everything else we do.

Parallelling Moore's law is the almost equally rapid increase in storage density, with the amount of data accessible on any square inch of media growing by a similar amount. These twin effects are contantly referenced in 'gee-whiz' articles in the computer press: "Why, just 7 minutes ago, a 14Mhz chip with a 22K disk cost eleventy-seven thousand dollars, and now look! 333 Mhz and a 9 gig drive for $39.95!"

All this breath-taking period doubling makes these measurements into a kind of physics of our world, where clock speeds and disk densities become our speed of light and our gravity - the boundaries that determine the behavior for everything else. All this is well and good for stand-alone computers, but once you network them, a funny thing happens on the way from the router: this version of the speed of light is exceeded, and from a most improbable quarter.

It isn't another engineering benchmark that is outstripping the work at Intel and IBM, its the thing that often gets shortest shrift in the world of computer science - the users of the network.

Chip speeds and disk densities may be doubling every 18 months, but network population is doubling roughly annually, half again as fast as either of those physical measurements. Network _traffic_, measured in packets, is doubling semi-annually (last year MAE-East, a major East Coast internet interconnect point) was seeing twice the load every 4 months, or an 8-fold annualized increase).

There have always been internal pressures for better, faster computers - weather modelling programs and 3-D rendering, to name just two, can always consume more speed, more RAM, more disk - but the Internet, and particularly the Web and its multi-media cousins of java applications and streaming media, present the first _external_ pressure on computers, where Moore's law simply can't keep up and will never catch up. The network can put more external pressure on individual computers than they handle, now and for the forseeable future.

IF YOU SUCCEED, YOU FAIL.

This leads to a curious situation on the Internet, where any new service risks the usual failure if there is not enough traffic, but also risks failure if there is too much traffic. In a literal update of Yogi Berra's complaint about a former favorite hang-out, "Nobody goes there anymore. Its too crowded", many of the Web sites covering the 1996 US Presidential election crashed on election night, the time when they would have been most valuable, because so many people thought they were a good idea. We might dub this the 'Icarus Effect' - fly too high and you crash.

What makes this 'Icarus Effect' more than just an engineering oversight is the relentless upward pressure on both population and traffic - given the same scenario in the 2000 election, computers will be roughly 8 times better equipped to handle the same traffic, but they will be asked to handle roughly 16 times the traffic. (More traffic than that even, much more, if the rise in number of users is accompanied by the same rise in time spent on the net by each user that we're seeing today.)

This is obviously an untenable situation - computing limits can't be allowed to force entrepreneurs and engineers to hope for only middling success, and yet everywhere I go, I see companies excercising caution whenever they are comtemplating making any moves which will increase traffic, even if that would be make for a better site or service.

FIRST, FIX THE PROBLEM. NEXT, EMBRACE FAILURE.

We know what happens when the need for computing power outstrips current technology - its a two-step process, which first beefs up the current offering by improving performance and fighting off failure, and then, when that line of development hits a wall (as it inevitably does), embracing the imperfection of individual parts and adopting parallel development to fill the gap.

Ten years ago, Wall St. had a similar problem to the Web today, except it wasn't web sites and traffic, it was data and disk failure. When you're moving trillions of dollars around the world in real time, a disk drive dying can be a catastrophic loss, and a backup that can get online 'in a few hours' does little to soften the blow. The first solution is to buy bigger and better disk drives, moving the Mean Time Between Failure from say, 10,000 hours to 30,000 hours. This is certainly better, but in the end, the result is simply spreading the pain of catastrophic failure over a longer average period of time. When the failure does come, it is the same catastrophe as before.

Even more disheartening, the price/performance curve is exponential, putting the necessary order-of-magnitude improvements out of reach. It would cost far more to go from 30K/hrs MTBF to 90K/hrs than it did to go from 10 to 30, and going from 90 to 270 would be unthinkably expensive.

Enter the RAID, the redundant array of inexpensive disks. Instead of hoping for the Platonic 'ideal disk', the RAID accepts that each disk is prone to failure, howsobeit rare, and simply groups them together in such a way that the failure of any one disk isn't catastrophic, because the other disks contain all of the failed disk's data in a matrix shared among the remaining disks. As long as a new working disk is put in place of the failed drive, the theoretical MTBF of a RAID made of ordinary disks, where two disks failed at precisely the same time, would be something like 900 million hours.

A similar path of development happened with the overtaking of the supercomputer by the parallel processor, where the increasingly baroque designs of single CPU supercomputers was facing the same uphill climb that building single reliable disks did, and where the notion of networking cheaper, slower CPUs proved a way out that bottleneck.

THE WEB HITS THE WALL.

I believe that with the Web we are now seeing the beginning of one of those uphill curves - there is no way that chip speed and storage density can keep up with exploding user base, and this problem will not abate in the forseeable future. Computers, individual computers, are now too small, slow and weak to handle the demand of a popular web site, and the current solution to the demands of user traffic - buy a bigger computer - are simply postponing the day when those solutions also fail.

What I can see in the outlines of in current web site development is what might be called a 'RAIS' strategy - redundant arrays of inexpensive servers. Just as RAIDs accept the inadequacy of any individual disk, a RAIS would accept that servers crash when overloaded, and that when you are facing 10% more traffic than you can handle, having to buy a much bigger and more expensive server is a lousy solution. RAIS architecture comes much closer to the necessary level of granularity for dealing with network traffic increases.

If you were to host a Web site on 10 Linux boxes instead of one big commercial Unix server, you could react to a 10% increase in traffic with 10% more server for 10% more money. Furthermore, one server dying would only inconvenience the users who were mid-request on that particular box, and they could restart their work on one of the remaining servers immediately. Contrast this with the current norm, a 100% failure for the full duration of a restart in cases where a site is served by a single server.

The initial RAISs are here in sites like C|NET and ESPN, where round-robin DNS configurations spread the load across multiple boxes. However, these solutions are just the beginning - their version of redundancy is often simply to mirror copies of the Web server. A true RAIS architecture will spread not only versions of the site, but will also spread functionality: images, a huge part of network traffic, are 'read only' - a server or group of servers optimized to handle only images could be served from WORM drives and serve the most popular images from RAM. Incoming CGI data, on the other hand, can potentially be 'write only' simply recording information on removable medai which can be imported into a database at a later date, on another computer, and so on.

This kind of development will ultimately dissolve the notion of discrete net servers, and will lead to server networks, where an individual network address does not map to a physical computer but rather to a notional source of data. Requests to and from this IP address will actually be handled not by individual computers, whether singly or grouped into clusters of mirroring machines, but by a single-address network, a kind of ecosystem of networked processors, disks and other devices, each optimized for handling certain aspects of the request - database lookups, image serving, redirects, etc. Think of the part of the site that handles database requests as an organ, specialized to its particular task, rather than as a seperate organism pressed into that particular service.

THE CHILD IS FATHER TO THE MAN

It has long been observed that in the early days of ARPANet, packet switching started out by piggy-backing on the circuit-switched network, only to overtake it in total traffic, which will happen this year, and almost certainly to subsume it completely within a decade. I beleive a similar process is happening to computers themselves: the Internet is the first place where we can see that cumulative user need outstrips the power of individual computers, even taking Moore's law into account, but it will not be the last. In the early days, computers were turned into networks, with the cumulative power of the net rising with the number of computers added to it.

In a situation similar to the packet/circuit dichotomy, I believe that we are witnessing another such tipping point, where networks are brought into individual computers, where all computing resources, whether cycles, RAM, storage, whatever, are mediated by a network instead of being bundled into discrete boxes. This may have been the decade where the network was the computer, but in the next decade the computer will be the network, and so will everything else.