On Fri, Apr 09, 2004 at 08:29:27PM +0100, Jim Dixon wrote:
Yes. I know what a tree is, and I am quite familiar with structure of the Internet. These very pretty pictures certainly look like the Internet I am familiar with, but don't resemble trees.
There's a continuum between a tree and a high-dimensional grid/mesh/lattice. A high dimensional grid pressed upon a flat surface looks like a bush at each node, with decaying connectivity density with the density. I don't have a sketch at hand, unfortunately, so I can't put both extreme cases next to each other. The Internet is a tree, not a mesh.
Over the last 30 years or so, various people have hypothesized about what the "killer requirement" might be. To the best of my knowledge, all have been wrong.
Computational physics in a relativistic universe imposes very clear constraints. There's nothing whatsoever hypothetical about these constraints. There's nothing whatsoever hypothetical that if you want to do relativistic cut-through switching of serial signals (10 GBit Cu, TBit fiber, LoS laser) you have to make a routing decision very, very early. Before the packet header has left the delay line (and the rest of the packet is streaming at you through the medium, whether vacuum or glass, or solid-state optical delay line). Medium is a natural FIFO, which of course only becomes apparent beyond GBit/s data rates. A photonically switched crossbar making use of it gets rid of expensive (time, energy) photon-electron-photon conversion, and scarce resource memory real estate, and memory bandwidth. It doesn't matter that in the current postdotbomb dark fiber is plentiful, and photonics is notoriously cash-drained. Traffic is growing, and will absorb those overcapacities eventually. Intermachine traffic, people will have become irrelevant pretty soon, if they haven't already (but for multimedia streams, which will saturate as well, because population grows way too slow to become visible, only technology deployment rate is visible). This is the reason why the future asks for specific frame/packet header layout, specific wiring of connections, and purely local-knowledge routing (extreme localization or elimination of admin traffic), with a routing decision done in ~ns domain (and below).
The Internet is quite obviously optimizing along certain lines. However, these lines don't follow any geographical geodesic, which was my point.
I'm not going to argue with you. Look up physical plots of connectivity over Earth surface. Start with GEO/LEO satellite, sea cable, then progress to large scale cable layouts, then to grassroot scale (city and neighbourhood and house level).
If you try to replace observations with theories, the most important thing is to verify that your theory corresponds with reality right now. If your theories aren't correct "currently", it is very unlikely that they will be a better fit tomorrow.
Non sequitur.
It isn't a minor point that the Internet is fractal. This is in fact what is consistent everywhere and has been, to the best of my knowledge, throughout the history of the Internet. If you go back to your pretty pictures and look, you will see fractal structures.
Dude, hypergrids *are* fractal. Not that it has to do anything with the current topology.
A geodesic is a minimal path in whatever geometry you are talking about.
The geometry on Earth surface is anything but whatever. Way above, with nodes in mutual plain view, it's plain old Einstein-Minkowski (basically Euclidian, with relativistic corrections).
If you looked carefully at traffic between European countries around 1999, it turned out that the minimal cost path between say German and France was in fact through Virginia. Traffic was following a geodesic -- but not a geographic geodesic.
Again, how about traffic in US? EU is weird, Asia is yet too new (but adapting very rapidly). Again, how about traffic in your above constellation in 2004? Again, how about physical cable connecting the nodes? I'm claiming peering arrangement evolve to make optimal use of given physical cabling. This is quick. On the longer term, physical and virtual (radio, laser) cabling evolves to minimize the load on existing links. This is slower, peering arrangements change in realtime in comparison, very like Franck-Condon principle.
As I recall, a 2 Mbps E1 between most major European cities and Virginia was about $30,000 a month, but an E1 across the English Channel was around $45,000 a month - 50% more to go 30 miles than to go 6,000. We had customers in Northern Ireland whose traffic to Dublin went first to London, then to our PoP in California, then to Virginia, and from there back to Ireland. This was our financial geodesic.
Why do people lay fiber in a specific place? How do peering arrangements evolve over time? How is the rate of optimization going to change if agoric load levelling is implemented at protocol level?
Indeed. But the point is that things tend _not_ to be optimized at the macro level; what happens is the opposite, micro-optimization around the results of previous decisions (some of which will have been just plain wrong). Roman engineers built roads a couple of thousand years ago, optimizing things according to then-current theories and strategies. We lay down rivers of fiber along those roads, reenforcing those ancient decisions, because the cost of reversing those ancient decisions, and all of the incalculable number of micro-decisions that followed, would be truly enormous.
You can see the same pattern working itself out now. A group of Japanese banks invested in a building in Docklands, Telehouse, to act as a backup facility in case of a disaster in the City of London. This turned out to be a loser, in financial terms. The Japanese had misjudged the market demand for this kind of facility.
Some telcos had put a few racks in the building. The first UK ISPs followed them there, because the facility was cheap. More ISPs followed. Some decided to build an exchange point there, the LINX, following somewhat misunderstood US models. Things mushroomed; the building, which had been quiet and empty, rapidly filled up with racks. The owners built another building across the street; investors built competing facilities a short distance away, to be close to the action. All of these were interconnected with more and more fiber.
Very interesting. Thanks for this story from the trenches.
The end result is that most UK Internet traffic, and a large part of European traffic, passes through what used to be a more or less derelict area of East London, all because of a planning error on the part of some Tokyo-based banks.
A nexus is a classical tree artifact. Once the network progresses along a meshed grid hugging Earth surface, we're going to see an increase in crosslinks and exchange points, crosslinking the branches.
Not at all. Everywhere we see the same pattern of pearl-like growth: someone makes a decision, and those that follow build around the first decision, micro-optimizing as they go along, creating the odd fractal shapes that are all around us.
future, everywhere, looks even more different. We're at the beginning of the optimization process. You can't cheat
I'm stuck with a notbook display and keyboard right now, but it would be fun to pull up graphs linking topology to geography, and change in such graphs over time. Informally, I hear topology converging to geography, but it would be nice to see actual animations showing it happen. physics
in a relativistic universe, in an economic/evolutionary context.
This isn't physics. It's much more like biology.
Biology is subject to physics. At all scales. When it comes to communication, constraints on energy and signalling and shape become especially obvious. Computational physics is just such a source of constraints. Except, here energy is not (yet) a constraint (dissipation rate is already), relativistic singnalling is (in comparison to biology, which is energy constrained, aand is waaay to slow to be relativistically constrained, only saltatoric spike propagation, which is arond 100 m/s, so it has to deal with latency as well when laying out the computational circuits). -- Eugen* Leitl <a href="http://leitl.org">leitl</a> ______________________________________________________________ ICBM: 48.07078, 11.61144 http://www.leitl.org 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE http://moleculardevices.org http://nanomachines.net [demime 1.01d removed an attachment of type application/pgp-signature]