Verizon DSL question

[mcmanus at ducksong.com (Patrick R. McManus)]

[John Jannotti: Tue, Apr 30, 2002 at 11:47:06PM -0400]

> frames have to get through instead of one in order to get a packet through.
> So, roughly speaking a 1% (Ethernet frame) loss rate becomes a 2% (IP

the truth is somewhere in between.

packet size is *not* independent of the packet loss rate. Smaller
packets are lost less often, though not strictly proportionately less often.

Also interesting, consider the resend granularity. If you lose 1 of
two packets you resend that lost packet. If you lose 1 of 2 fragments
you have to resend them both.

> Matching your IP packet size to the true MTU size of the path gets you down
> to the true loss rate, and TCP works better.

yep.. PMTUd exists exactly because fragmentation hurts so bad
(fragmentation doesn't even exist in ipv6 - PMTUd is a
requirement). but PMTUd has significant costs too, much better to
short circuit the whole process by setting your host's mtu
appropriately when you know there is a smaller mtu in the path (this
also serves as a hint to the other end to use your mtu, the PMTU cost
is much more expensive to him because the fragmentation point is so
much farther away from his perspective.).

-P