[PLUG] Bandwidth allocation: was Torrent download speeds

Keith Lofstrom keithl at kl-ic.com
Mon Jan 14 01:00:56 UTC 2008


Oh boy, another stale subject line to retarget.  I may have found my
purpose in life!

On Sun, Jan 13, 2008 at 03:42:46PM -0800, Mike Connors wrote:
> Any business internet account would get full bandwidth accounting 
> complete with bar graphs, pie charts, and threshold alerts for 
> over-usage. The point of it is to see how much actual bandwidth you're 
> consuming. I would imagine the average 40hr/wk working person is on the 
> internet from their home approximately 20 hrs/wk. And for that same 
> average person a very small percentage of that time is for 
> high-bandwidth usage.
> 
> If I'm hogging bandwidth and making internet access in my neighborhood 
> miserable for everyone else than I should pay more. You have to 
> understand how these companies dole out bandwidth. The industry standard 
> is to over-subscribe 4 to 1. They are just trying to get as many people 
> signed up for access and are banking on the fact that *most* people only 
> use their connection a few hrs/day and *generally* don't consume large 
> amounts of bandwidth for long periods of time.

Agree with most, though I suspect the "average" person (besides having
one mammary, one testicle, and being mostly dead) is on the 'net far
less than 20 hours per week.  The average person is also not using the
telephones, the roads, or the sewers continuously either, which are all
systems that work only because of the averaging of large numbers of
random and low-usage customers.

In the early nineties, I designed a line of chips called "non-blocking
crossbar switches" for a company called I-Cube Design Systems.  They
were originally intended for hardware logic simulation, but I-Cube's
original simulator customer ended up in legal battles over patents.
I-Cube found themselves a new customer that could use switches, a small
Bay area outfit called Cisco Systems.  The 'net took off, Cisco took
off, and I-Cube had by far the best switch chips at the time.

Most companies provided switches that were small and complicated and
slow.  They tended to use something called "Banyan Trees" based on the
idea that the average amount of traffic needed fewer switches than a
complete "any to any" matrix.  That made the switches "blocking", that
is, an established route could prevent another from taking place.  This
only occured in periods of high usage.  However, usage of any 
oversubscribed resource is usually "fractal", meaning that there are
periods of very high usage that don't conform to bell curve statistics.
For land-line telephones, it is Mothers Day.  For sewers, it is the 
Superbowl station breaks.  And for the internet, it is news events and
(locally) slashdotting.  Banyan tree networks crumble under the load,
way-overbuilt non-blocking switches do not.

The I-Cube switches were any-to-any and non-blocking - you could light
up the entire switch without blocking messages.  They were also easy to
test, cheap, low latency, and wicked fast.  I-Cube was growing 30% a
quarter, with 50% profit on sales, when the V.C.s decided that they
couldn't explain to IPO investors what the company did, and sent I-Cube
off chasing negative-profit easy-to-explain stuff like Ethernet chipsets.  

The next big shakeup in the net will be the next big event that sends
the zillions of "average" users to camp out online, looking at videos
they can't get from the network.  An oversubscribed network may fail,
angering the "occasional" customers and making them easy pickings for
competitors after the Big Event is over.  

I suspect a lot of customers have 'net connections so it will be there
in an emergency, and are not big everyday users.  Whichever network is
best "shaped" for extreme usage periods will pick up a lot of customers
after the next crisis.  As the cost of provisioning bandwidth continues
to plummet, extra bandwidth, and the ability to allocate that bandwidth
to the maximum number of "average" customers, is what forward-thinking
ISPs should be looking for.   A good proxy for that might actually be
open-source downloads and bittorrent - as long as you can cut throttle
those down when the fecal matter hits the air circulator.

Keith

-- 
Keith Lofstrom          keithl at keithl.com         Voice (503)-520-1993
KLIC --- Keith Lofstrom Integrated Circuits --- "Your Ideas in Silicon"
Design Contracting in Bipolar and CMOS - Analog, Digital, and Scan ICs



More information about the PLUG mailing list