[NCLUG] DSL Throttling or General Congestion?

[John L. Bass]

About 10 years ago I wrote a rather scathing summary about why ATM is
horribly bad for the internet, and it's cell dropping response to queue
overload is why.

When you drop 10 packets to create throttling, create some degree of
additional work, offset by clients backing off. The rest of the packets
are delivered normally.

Droping 10 cells, corrupts up to 10 packets, without lowering the down
stream work significantly. And yes, 10 clients will backoff -- maybe
if they are TCP.

The problem for both systems is that increased streaming UDP doesn't
back off, so the primary load sources will continue to congest overloaded
network points, and totally trash the TCP users access to the internet.

IP based congestion decisions can make wise choices about which packets
to trash, ATM cell systems are greatly hampered, as a single ATM cell
doesn't contain any information about packet type.

Packet loss as a throttling solution is just absolutely wrong when you
follow this line of thought to conclusion ... it doesn't work, and totally
trashes the wrong users.

THere are ways ATM could fix it, by sending congestion information to the
edge routers in the cloud, but that isn't part of the protocol. Frankly,
memory is cheap these days, and ATM is just the wrong solution. It was
the wrong solution 10 years ago.

John