[e2e] Skype and congestion collapse.
Clark Gaylord
gaylord at dirtcheapemail.com
Thu Mar 10 21:38:15 PST 2005
Alex Cannara wrote:
> Syed, the misconception is that apps are to manage network resources.
> That's not how more robust systems, like the established telco and
> private network systems work. Systems must protect themselves, even a
> car's gearbox, from external abuse. The Internet is only different in
> that attention wasn't paid to doing that in its design.
But that doesn't mean that we can't pay more attention to it now.
Gearboxes didn't protect themselves originally either, but synchromesh
and fluid drive have done a lot to address that. The advances in
computing are due to algorithms, and in recent years that has been
parallel algorithms. That's what TCP is: a distributed, parallel
algorithm that does a reasonable job of equitably sharing limited
resources. But you are correct that our network systems can do more to
enforce fairness and compliance at some level.
By protecting TCP traffic from non-TCP-friendly traffic, by ensuring
that TCP plays by the rules, by judiciously and scalably employing
policing and differential queueing -- we can build a much better
system. But what makes the Internet a superior *engineering* solution
is that it doesn't try to over-specify everything. "Do just enough" is
what makes this work, and it isn't due to laziness: it is due to the
inherently superior scalability of this approach. We don't need to
debate whether the Internet is a superior solution to the PSTN; reality
has already proven that. What is left is for those of us who wistfully
harken to when TDM was king and everyone had a 3270 to understand that
we need to crack our bell-shaped heads open and get up to date. Poisson
and Erlang don't apply anymore, but doesn't mean we can't model (or that
we can't extend them in interesting ways, btw). Call admission doesn't
apply anymore, but that doesn't mean we have to drop all packets equally.
"What do we need to do to make the Internet better?" is the question we
need to be asking.
--ckg
More information about the end2end-interest
mailing list