cost of bit-pipes

[David Farber <dave () farber net>]

Begin forwarded message:

From: mo () ccr org (Mike O'Dell)
Date: July 24, 2007 4:02:12 PM EDT
To: dave () farber net
Subject: cost of bit-pipes



Just the other day, I read an adage that I wish I could take credit for,
but for the life of me cannot find the correct attribution. (A  
pointer would
be most appreciated.) here it is anyway:

        First Rule of Networks:
                Pipes can be made wider but not shorter.

the actual *cost* to build, own, and operate a transmission
facility is well-predicted by the PRODUCT of SPEED and DISTANCE.
There is an additional constant of proportionality which arises from the
interaction of transmission technology and a highly dis-continuous
cost/distance function, but to one digit of precision, NETWORK
CAPACITY, meaning the ability for a network to move bits a
non-trivial distance, is measured in (Gigabits/second)*(kilometers).
Note that "raw bandwidth" is routinely measured in just
Gigabits/second.

        THE UNITS ARE NOT THE SAME!

it is indeed the case that within particular bands of scaling,
the cost contribution of the distance term is sufficiently
uniform (or negligible) that it's normalized away in casual
conversation. (Convenient, but sloppiness that ultimately
results in the confusion manifested in this discussion.)

A good example is an Ethernet patch cord. Longer ones are more
expensive than short ones, but the difference is small enough
that nobody seems to care even when comparing Ethernet patch
cords with building runs; the distance term is set equal to 1
and comparisions are based on "speed".  (NB - if you are
concerned about *latency*, these differences actually matter
and people engineering for, etc, cluster MPI performance are
much more attentive to getting the details right.)

If, however, you wish to send a gigabit/second several decimal
orders of magnitude further than an in-building Ethernet drop,
the distance term suddenly matters dramatically. There are
multiple reasons, but an important one is that same electrical
level signalling that gets you around a building does not work
getting you around a state or country, even if the bit rates
are the same. (Again, students of delay and latency are painfully
aware of the miserably slow speed of light and are already
writhing in pain.)

and when the distances get even bigger, or possibly have to
transit an ocean, the distance-related term can quickly grow
to monopolize the other terms.  having to install equipment
every 100 KM on a transcontinental fiber span is a non-trivial
cost contribution all by itself, regardless of how fat the
pipe might be along the right-of-way. and no amount of
spectacular modulation technology can alter how long it takes
a bit to get from one end of the fiber span to the other.
Complain to Prof. Einstein, not me.

at the end of the day, there is no particularly effective
way to move San Francisco closer to New York City, and that
distance matters a very great deal, and will always matter
given sufficient scale.

(Note: statistical multipliplexing is another matter entirely
and should not be entangled with the realities of facilities cost.)

        cheers,
        -mo



-------------------------------------------
Archives: http://v2.listbox.com/member/archive/247/=now
RSS Feed: http://v2.listbox.com/member/archive/rss/247/
Powered by Listbox: http://www.listbox.com