From: Eugene Leitl (eugene.leitl@lrz.uni-muenchen.de)
Date: Sun Dec 26 1999 - 22:58:44 MST
(Inadvertedly, you have mentioned one of my favourite subjects. I went
once crazy (crazy like in not sleeping for a few days and thinking
hard enough to fear you're starting to crack up -- in other words NOT
my normal beerguzzling groundsloth behaviour) for it for a while, a
few years back. A classical case of the tail shaking the dog. I will
rant (this is a threat, in case you're wondering), and most of it will
not appear technical. Trust me, deep down below it is).
There is a problem with all the networking technology we're now
using. It operates in the physical reality with all their constraints,
but gadgets nor protocols are aware of it. Because of this there must
be eventually a hard break when you have to throw out everything and
start from scratch. A radical redesign. (Everybody just loves
them. There is gotta be a reason for all the braindead specs out
there). Of course everybody will say you're a nutcake if you tell that
them that today.
Networking started when hardware was exotic (=scarce), and spatially
separated by lots of prairie in between. It was hard to see in advance
for anyone that this would change dramatically at some point. People
are naturally biased towards sequentiality and centralism (which is
quite a sad joke, if we think of how our very own wetware is organized
at the deeper level). So they gave nodes unique numbers and thought
each node could have enough knowledge about the whole network. The
connectivity was not a mesh, it was a tree at best. It still mostly
is. This works for quite a great while, and still does, in a
fashion. Unfortunately, it starts breaking down when the nodes
(literally) cover the planet, and start expanding into space as a
relativistic bubble.
Computational physics as we know it makes local links faster than
nonlocal links. Faster in terms of latency (relativistic lag) and in
terms of bandwidth, regardless of the technology used. (Barring
generic lunatic fringe things like spacetime plumbing and FTL
signalling; let's worry about them when we have to. Foresight has to
have limits somewhere). On the scales we're aware of, space itself is
not tagged (is rotationally and translationally invariant) unless you
put matter in there which is not. Assuming (a gedanken, today) you
fill (relatively flat) spacetime with random distribution of nodes,
certain things immediately spring to mind.
The network is kinda huge. (Duh!). Fortunately, combinatorics with a
relatively few bits can make you address a lot of things. Like,
really, a lot. Let's assume the nodes are already labelled (how this
came to be, I'll address later). The node distribution is largely
homogenous, and connectivity is essentially relatively low and
local. Line of sight, in other words. Twinkle, twinkle little node.
You can't know the entire connectivity of the network, because you'd
waste too much bits to represent it and would have to think too much
on how to get theah from heah. Relativistic lag also applies to
molecular switches, since they switch on the order of magnitude it
takes light to traverse the space they're in. You don't have to have
global network knowledge, provided there is an orderly way in how the
node tag is varying while you're travelling through space. IDs must
span a coordinate system (who woulda thunkit?!). Things which talk to
each others a lot, and frequently, be better in the neighbourhood, the
nearer, the better.
A message in transit is smeared over space. It takes time for the bits
to fly past you, and the head arrives before the tail. Because the
nodes are aligned on noisy grid and their IDs span a coordinate
system, we don't have listen to the entire message and to store it if
it says where it wants to go right from the start. You (cut-through)
switch, not route. Sometimes things go wrong, and you have to store-
and-forward if you want to signal reliably across the whole of the
network, so the message better be short. Takes shorter to process, and
fewer bits to represent. When switching, you look at the message in a
very short FIFO, long enough to look at the header. You switch each
message, but also clone it into your message cache until you've gotten
an ACK for a few nodes in your local sphere so you're sure it was
passed on correctly even if one or a few nodes are going to die
simultaneously. After the passed on message is thought to be stale
enough, it decays from cache (built-in bit rot). Long-haul messages
(geometric distance from own and target ID) should have higher
priority, since retransfer can take forever, affecting every relais
inbetween. Otherwise, you have to resend it. After a number of
unsuccessfull tries you have to give up, since your cache is not
infinite, and every bugger in the local supercluster is talking a blue
streak.
The protocol has to be homogenous. Computronium cells a few microns
away and nodes in Magellan cloud should be talking in the same
language. On the other hand, if only neighbours talk to each other
there is no need to engage the high order bits. Binary representation
defines a natural hierarchy, so let's use this fact, and save space.
If our network has grown from a nucleus (as far as we know current
interstellar dirt is dumb as... dirt), there is no trouble to find an
origin. The growing front labels itself as it grows, by looking
back. At some point, hopefully as late as possible, it will run smack
into an alien domain. Ouch. The addressing there is incompatible,
unless spacetime has a truly global unique arrow hidden in there
somewhere, which currently appears doubtful. Due to relativistic lag
it will take forever for one domain to overwhelm the other (akin to
spiralling demon waves), so this is not practical. Maybe frontier
nodes belonging to several domains can do something here. Maybe
not. Did I already mention braindead specs?
Domains will also be created by spontaneous symmetry breaking from
noise. As above, phase change of tagging will take forever to
propagate, so effectively the total whole is multidomained of roughly
same grain size, if seen from a large distance. Very large distance.
Enough PostSingularity stuff. It will all be totally different,
anyway.
So, let's get down to earth. Earth surface, to be precise. Time: near
future. A decade from now. Less? The number of smart gadgets have
multiplied sufficiently so that there is a nonnegligeable number of
them in a given patch of a civilized world. Cell phones and base
stations, roofnet nodes, PDAs, cars, street signs, wristwatches,
toasters, blimps, drones, sats, planes and other sundry airborne
vermin. Some of them are mobile. Most local networking is wireless,
using digital pulse radio and optical. If the gadget soup is thick
enough, it creates a network by virtue of being in the area. Provided
most nodes are immobile, or at least have a fair inkling of their
bearings, you can pluck your position of of thin aether, by mutual
triangulation. Realtime GPS++, cm resolution or better. Since
everybody is mutually triangulating like crazed weasels anyway, why
not putting data cargo in all the pulse sequences?
Even more down to Earth. Time: near present. (Do you like it, crawling
here in the muck? Thought so). The ISPs have wisened up somewhat, and
don't route your packets to the neighbour next door via
Australia. (They now do it via Seychelles). They mostly use
cut-through switches, because streaming multimedia has finally hit the
streets, and grandma doesn't like second lags in audio when talking to
little Johnny. Now latency metric gotten from pinging "nearby" IPs
doesn't give you just garbage anymore. GPS in cellphones make once
quaint icbmto://N34.06'03.1"W117.34'12.7" pseudotags semi
ubiquitous. Trimble-toting technicians started flashing WGS 84
position fixes along with assigning IPng addresses, since otherwise
hardware can be hard to find (particularly, for semiblind robot
troubleshooters, in the prototype stage). Finally, finally!, web
spiders have started picking up position information. Marketeers
suddenly find fresh cookies even more irresistible.
Er, where did I want to go today? Did I mention beer, already?
Michael Nielsen writes:
> (Inspired in part by Eugene's posting of Phil Agre's newsletter to
> >H-tech.)
>
> The internet is a terrific way to connect people with common interests,
> but it's a little lacking in that the contact usually remains online,
> except in the happy coincidence that you discover a kindred spirit in
> close geographic proximity. Meanwhile, in your actual physical
> neighbourhood there are doubtless many people with whom you'd like to make
> contact, but who remain unknown to you.
>
> What techniques can be developed to connect people who (a) live in close
> proximity, and (b) have common interests?
>
> One simple solution that comes to mind is smart online map or telephone
> directory services that not only understand real-world geography, but
> which are also clever enough to search out personal web pages with
> real-world addresses attached. Use the online service to find the
> street you live on, and discover that three blocks up there is a person
> with interests in Cajun cooking, parallel computing, transhumanism, and
> Jackie Chan movies...
>
> Anybody have other suggestions for connecting real-world and online
> geography?
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:06:12 MST