[p2p-hackers] MTU in the real world

[Matthew Kaufman]

Nick Johnson:
> Not exactly. Statistically, it'll taper off asympotically. 
> For example, 10% actual packet loss with each packet 
> fragmented into 10 is (1 -
> .9^10) or 65% apparrent packet loss.

Actually my claim is correct "if... 1 in 10 fragments lost causes every
packet to be missing at least one fragment" (admittedly not the likely
case)... However you are absolutely right that, statistically, the most
likely case is that it will cause the packet loss you've described. Still a
lot worse than the original 10%... Related closely to why congested ATM
networks cause IP to experience a "brick wall" at the congestion point.

Short story: fragmentation is bad in almost every case. Unless you are in
one of the very rare and very specific cases where it makes sense (an
example might be where you can send giant packets efficiently, and can
afford to have them fragmented elsewhere, and can manage the extra loss risk
(and as Serguei pointed out, receiving host overhead when it occurs)), you
should be trying to stay under the path MTU, either by discovering the
actual MTU, or making your best guess as to what it might be.

Matthew Kaufman
matthew at matthew.at
www.amicima.com