Neural Net Routers To Speed Up Net

Virek writes "Researchers in the UK have developed a prototype packet router that is potentially much faster than existing digital technology. They are using a combination of a neural network and "diffractive optical components" (er..?). If they are successful in developing a commercial product they predict that this could significantly speed up the net. "

Comments

Sounds |\|337, but it's not a router.

[Phizzy]

Routers have two components : a routing component and a switching component. The switching component takes input on an interface, moves it across the backplane to the routing component, and then and pushes it out another interface when the routing component is done with it. The routing component reads the destination from the IP header, compares it to it's routing table and decides whether it is routable, and if so, which interface it will go out. It seems to me that this technology would only take care of the level 2 (switching) layer of networking, which is not the major area of concern in networking right now.. With existing electrcal technology, we have backplanes that can push 160Gb/s of IP traffic. That, on one device, is enough to push 16 OC-192s. The problem is if a single device had to push 16 OC-192s at wire-rate (full speed, basically), the route processor would not be able to handle the IP lookups to feed the backplane. So this technology (if they can get it to work.. which it seems like they should be able to) _will_ be useful once we get a route processor that can keep up with it.

This tech looks like it could be separated from a router to be used in an ATM switch, but by the time it is viable, we probably will be using MPLS and POS more than ATM, but there will always be a need for fast layer 2 technologies.

I like the current trend in networking to move away from electrical processes in switching.. once we can get a fully fibre-based system we will be able to remove much of the inherent latency in telecommunications.. like the 100ms lag to go across the atlantic, or the 60ms lag to go across the United States. Once we have comparatively few electrical devices taking care of the routing and all optical devices doing the switching, the only thing that will cause lag will be processing and the speed of light.

//Phizzy

Oh frick, diagnostic nightmare

[Morgaine]

One thing that distinguishes neural networks from other approaches is that the "cleverness" (ie. the effective rulebase that determines what happens) is embedded implicitly in the internal state of the network as a set of trigger weightings, and the pattern of weightings has absolutely *ZERO* obvious relationship with the work being performed in any sense that is meaningful to humans.

It's bad enough trying to diagnose routing problems right now, in a world of dynamic but still fairly deterministic routing algorithms. Add neural networks into the equation and all bets are off: we'll never be able to determine why any particular route was taken. (An explanation of "because those 57 nodes fired" is not particularly helpful).

Oh joy.