[PLUG] Comcast, Verizon FIOS, QOS

[Keith Lofstrom]

One does not have to hypothesize malice when indifference will explain
the results.  To make voice over IP work properly, you don't need a
high packet rate, but you do need low delays for guaranteed packet
delivery.  Packets can't be delayed very long or dropped too often -
they must travel through the system with high QOS or quality of service. 

High QOS is expensive in a bursty traffic infrastructure like the
internet and the backbone.  To make sure high QOS packets always
(or even mostly) arrive on time, you must dedicate part of the fixed
channel and switch bandwidth to it, identify high QOS packets, and
route them through this set-aside bandwidth, and limit "Mother's Day"
initiation of connections.  Otherwise, the inevitable and momentary
burst peaks of internet usage will saturate your pipe and the high
QOS packets will arrive too late. 

This provisioning must occur at every point along the way, and must
be paid for in advance, so it is there during "burst times".  ISPs
must purchase high QOS from their carrier, and dedicate and program
switch hardware, and so forth.  I imagine they do this for clients
who pay extra for the arrangement (say, companies that route their
corporate phone and video systems through the net) but ordinary
residential users have not paid for this and won't get it.

I don't use VOIP, but I see the burstiness when I look at Youtube.
Often, the content delivery bar crawls along just ahead of the viewing
bar, and the video stops a few times per minute while more content is
buffered.  Sometimes, I just pause the video near the start and wait
for a bunch to be delivered, and for popular stuff (remember
Dr. Horrible?) I would either use "youtube download" apps or just
pause for a half an hour before watching the 15 minute video.

The internet is oversubscribed.  If all of us took our 5MB download
links and opened the throttle at once, you betcha the result would be
that data would come dripping out for all of us, because the cable
(Comcast) or the switches (Verizon) would saturate.  It costs far too
much to buy enough cable and switch so that everyone can have all the
bandwidth, all the time.  And it wouldn't trickle, it would drip,
because they don't pay attention to individual subscribers three
switches up where the saturation may be occuring.

Comcast especially - they have very definite limits about how much data
they can push through a cable repeater amplifier without frying it, and
when they increased their maximum download rate for consumers, all they
really did was took the same limited bandwidth and increased the
oversubscription ratio - individual users can take a bigger slice of
the limited bandwidth, and saturation occurs much more often.  Third
party services that were designed to tolerate only a low amount of
saturation (like VOIP and realtime streaming video) get hammered.  

And why should these ISPs that offer phone service do anything to fix
that?  As long as their minute-by-minute usage averages out to the
advertised bandwidth, they've met their obligation to their web-surfing
and downloading customers, whereas increasing their millisecond-level
QOS for all internet traffic requires a hefty investment in cable and
switches and payments for backhaul and intertie with other services. 
What they get for that investment would be less phone business.  
If the network never saturates, they are paying too much money for
it, and charging the customers too much to pay for that.

I expect that some ISPs may intentionally drop some tcp packets that
seem to be VOIP related (requiring a delaying retry), but I doubt they
have enough buffering in the system to delay them. 

I can also imagine VOIP protocols that send multiple copies of the same
sound packets, or overlapping packets, at slightly different times,
and that might break up a little less.  However, too much "oversending"
would just add to saturation, if VOIP ever became popular enough to
use a significant of the net.

So we are left with the sad truth that the internet is a packet network
subject to economic limitations affecting packet delay and resend, that
expensive provisions must be made to provide the minimal QOS needed for
VOIP, and if you aren't paying for low latency and high QOS, you will be
disappointed if you are expecting it.  You want a phone?  Pay for phone-
quality service.  You want a telegraph?  The internet is your baby.

My solution for The People's Network?  First, fiber to the home, 
paid for by the consumer, but to first level switches that can connect
to multiple ISPs.  Second, local mirrors and multicast to lower the
burden on top level switches.  Third?  Track down spammers and other
negabit providers and sell them for dog food.  Fourth, real time 
competitive pricing for the packets, with high prices for high QOS and
very low prices for bulk transfers between mirrors.  Fifth, consumer-
owned agent software that navigates the maze of mirrors, providers, and
service levels, helping the bandwidth consumer balance costs and desires. 

The net effect might be that I can have an Ubuntu download in 15 seconds
for $10, 5 minutes for $1, and overnight for $0.01.  I can make a three
minute phone call to the house next door for $0.05, or from the bottom
of the Grand Canyon for $5 .  And my telephone will give me the price
estimate before I make the call (as well as an estimate of how pissed
I will be when I see the weekly bill aggregating all those decisions).  
Behind the scenes is a "super ebay" of bidding and selling which I don't
have to worry about.


My cred?  I designed silicon for Icube Design Systems, and those chips
switched much of the internet in the early 1990's for companies like
Cisco and others.  We sold "complete matrix" switches suitable for
bursty traffic, while the competition sold overly-baroque Banyan
tree and other switches more suitable for fixed channel telephone
switching.  We creamed the competition, until idiot VCs creamed us. 
While I am sure much has changed in a decade, the nature of the
traffic and the behavior of the switching investment (which goes up as
the number of endpoints squared, and down with Moores Law) has not.

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