Home media servers, AUPs, and upstream bandwidth utilization.
scg at gibbard.org
Wed Dec 27 21:09:28 UTC 2006
On Sun, 24 Dec 2006, Roland Dobbins wrote:
> In the U.S. and Canada, the expectation has been set to an assumption of
> 'unlimited' bandwidth consumption for a fixed price in the consumer market.
> AT&T WorldNet helped popularize that model early-on (you can thank or curse
> Tom Evslin for that, according to your inclinations, heh), and it has become
> de rigeur for most U.S./Canadian broadband SPs to follow suit.
> Which raises a related question of whether North American operators believe
> that offering value-added services such as placeshifting (which is a familiar
> enough concept that a significant population of the userbase seem to grasp
> the idea without a lot of explanation) might prove amenable to metered
IANAE (I am not an economist), but:
At the Peering BOF in Saint Louis, somebody commented on how when you put
a group of network engineers into a room, pretty soon they stop talking
about engineering and start talking about economics. There are a whole
bunch of factors that go into what it costs to provide a service, which
differentiate the costs of upstream versus downstream traffic:
In theory, downstream traffic uses more resources. Given hot potato
routing, traffic gets handed off from the sending to the receiving network
close to the source, and the receiving network pays the cost of hauling
the traffic over a long distance. If we could measure the cost of traffic
purely in bits per mile, upstream traffic would be cheap, and downstream
traffic would be expensive.
A lot of networks have ratio requirements in their peering policies. If
they're receiving more from a peer than they're sending, they threaten to
depeer. On the surface, this makes sense. They're receiving traffic and
having to pay to carry it, while the other network isn't paying as much to
hand off traffic locally. But it leaves content heavy could-be
competitors desperate for inbound traffic.
Exchange point ports and private peering cross connects tend to be
symmetric. If we ignore the longer distances that inbound traffic has to
be hauled, and the resulting increases in required capacity of inter-city
backbone links, we can say that the cost of handling inbound traffic has
already been covered in the course of handling outbound traffic. There
may be one too many "if we ignores" in the previous sentence, but there's
enough fiber in the ground between major US cities (as well as to Western
Europe and East Asia) that it's pretty affordable.
What's a network desperate for inbound traffic to do? They either risk
losing their peering, or they find customers with lots of traffic flowing
inbound. When there are lots of networks with this problem, they start
competing for the inbound-heavy customers. It's been a couple years since
I last shopped for inbound-only bandwidth, but at the time the cost for
inbound-only bandwidth in large quantities in major US cities was rapidly
approaching zero. This was putting downward pressure on what those who
already had balanced ratios could charge, as well.
Inbound bandwidth, while theoretically more resource intensive, was cheap
or free. To the extent that the costs weren't just resulting in more debt
for the transit providers, the outbound-heavy customers were subsidizing
this. This meant big price differentials between outbound and inbound
Then we've got the issues of what gets charged to the users:
In the US, downstream bandwidth doesn't cost much. End user customers are
expected to use small amounts of downstream bandwidth. Billing the end
users for the bandwidth consumption would cost more than covering the cost
of the occasional excessive user. In other words, it is "too cheap to
meter." Hosting is a different story. Hosting customers often use enough
bandwidth to affect their providers' transit bills, or even to require
infrastructure upgrades, and the traffic is almost all upstream, the
expensive direction. Hosting is almost always billed by usage.
In some other parts of the world, where the ISPs bear the cost of bringing
their inbound and outbound traffic over expensive connections from far
away places, the cost model can be quite different. As the costs of
transport and transit go up, the costs of dealing with excessive use
surpass the cost of metering. There are lots of places where it's common
to have separate rates for local and long-distance Internet, with the long
distance connectivity being billed per bit.
We keep being able to push more data through the same fiber, and more data
through equivalently priced hardware, than previously possible. But I
suspect that if traffic patterns were to change and cause the costs of
dealing with excessive use by US broadband customers to go up
considerably, we'd see the billing systems change to reflect that.
More information about the NANOG