On Mon, Feb 26, 2007 at 10:46:38AM +0100, Eugen Leitl wrote: I'm going to be terribly lame, and comment my own post. Perhaps it's going to make more sense now, perhaps not. If it doesn't, this conversation is perhaps not sufficiently intelligent, and I'll drop it.
> A lot of the original design for the protocols didn't address several > basic issues. > > - the network was small and irregular The early ARPAnet had few nodes, typically connected by long-distance links which were slow, using core for FIFOs. Today's WANs begin at 40 GBit/s and go up to PBit/s, soon more. There's a dense network coverage -- at least in urban areas. There are GEO and LEO satellite constellations. > - transfer speed was well below relativistic Photonic networks work at the speed of light in glass, so at high data rates a length of fibre is a FIFO. This is a lot like storing bits as circulating sound waves in mercury-filled tubes with ultrasonics transducers. At mere 1 GBps a bit is just 30 cm short -- in vacuum. A km gives you a 3 kBit FIFO, which is more than enough to contain the packet headers, and a bit of the packet. At TBps it's a cool 3 MBit/km, but it's getting crowded, because of the many colors. But very short pulses are multihued by nature, anyway. Isn't it fun when networking becomes indistinguishable from physics? > - logic was electronics, not photonics > > If you fire photons in a fiber FIFO, very much like sending > information by firing tracer bullets into the night, you want > to make the routing decision while the packet headers are still > streaming past you. It is a switch, not a router. Because of timing, Basically, you're treating you medium (vacuum, fiber) as FIFO, and use additional photonic delay lines (a length of fiber rolled up, a little cloud of BEC, whatever) to be able to process your header information and do your routing decision (on a forked copy of the original stream) while your data packet spends a little time in the waiting loop. > you need pretty shallow logic, since invididual gate delays add up. This means your algorithms need to be dead simple. There's absolutely no time for iterating things, you have to decided within a gate depth of some 10. This is not much of a handicap: people process complex stimuli within 50 ms, leaving only few neural processing cascades. But this is why you need to use a Gray-like coding (no time for singularities, where many bits in a register flip at once), where flips are small, and local. A natural surface for a gravitational assembly is the surface of the sphere, or at least a fair approximation thereof (a geoid -- mere Ethernet MAC address space allows you to put one machine/m^2 on Earth suface, or something), or an orbit. The one encoding that is right is that minimized the amount of bits flipping and limits the locality of the bit flipping if your node is in a (circular) orbit. As to the math, let somebody else do the math. > Worse, you don't want to convert photons to electrons and back, > so you preferrably would do this with photonic gates, which are expensive > in numbers and power dissipation (NLO usually appears at higher fluxes). There was a time, not too long ago, when a bit toppling out the rack would squash your foot. These times are not quite past in photonics, where a typical lab assembly easly fills a crowded optics table, or a couple. Integrating a few 10 kGates worth photonics upon the area size of you fingernail is yet in our future. > The idea is to make an astonishingly dumb network, one that is > basically only a thin decoration upon this universe's basic laws. > Once you fire your stream of tracer bullets, they go on own power, > until they hit a node somewhere, and have to be redirected on the > fly to a different sky segment. Our networks are too smart for their own good. You can skip on the smarts by wiring your network the right way, using small-world layouts, especially highly regular small-world layouts. It's still a lot of fiber, which is however laid at residential, or polis circuit level. When people are bits, you need one heck of a network crossection. Especially, since at just 10^6 speedup a wall clock day turns into three kiloyears, and you can locate objects by their light wavefront delays, just as we do with sound. > I've once experimented with node layout on global scale (where the > geometry is the surface of a sphere, or a concentric thick shell, and > you can actually build a local-knowledge only routing scheme which > is quick enough. Basically, you don't store global routes, but I don't have to point out that our current protocols have a severe routing table size growth problem, and that some 10-20% of the entire global node connectivity isn't really that global. There is dark matter on the Internet. And it's probably growing. > deviations of the local lattice defects from perfection. The node > address is a binary string representing the current node position > (something like a polar coordinate system, but with a Gray-like Polar, because gravity likes things curved (yes, not really, Einstein, I know). > encoding with local bit flips and no addressing singularities). Addressing singularities when crossing imaginary boundaries can be Bad News. Nothing sucks more like completely losing flight control at multiple Mach. > I was thinking about using a hierarchical addressing scheme to > switch to individual gravitational assemblies (Earth/Moon, Sun, > galactic system, etc). Birds revolve about Earth which revolves about our local star, which revolves around local galaxy center, which form clusters and superclusters, and so ad infinitum. You can address all of it with a very modest number of bits, though, of course, a MAC won't be enough. -- Eugen* Leitl <a href="http://leitl.org";>leitl</a> http://leitl.org ______________________________________________________________ ICBM: 48.07100, 11.36820 http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE
