Port numbers and traffic data

[Ian Batten]

On 24 Apr 2008, at 15:19, Tom Thomson wrote:

> 23 April 2008 13:39 Ian Brown:-
>
>> IP works just fine over unreliable link layers. Reliability comes
>> from the layer above (TCP) when required. Clearly IP works less
>> efficiently at very high utilisation rates on link layers with frame
>> sizes below typical packet sizes. Therein lies the solution ;)
>
> If you are saying that the solution is to have local networks which
> either are very reliable or have frame sizes as large as typical  
> packet
> sizes, that certainly works - it is one mechanism for achieving the
> actual requirement (that if

One problem with IP over ATM --- in any form, be it LANE or Classical  
--- is that it completely bypasses Path MTU discovery.  The rest of  
your posting makes clear the reasons why it's a catastrophe, but it's  
worth pausing to think of one of the major changes to IP as deployed  
over the past twenty years.  And that's the move from the default MTU  
for off-net traffic being 576 bytes to the current position where its  
assumed to be at least a standard ethernet frame.  That happened  
precisely because backbones handled ethernet-sized frames in their  
entirety, so it was sensible to work in frames of that size: Path MTU  
discovery is an implicit admission that 1500+ is the default case, so  
you handle the exceptions by exception.

But ATM transport breaks all that, especially under load, and the only  
fix is random early drop: an ATM switch whose upstream is under stress  
should start dropping entire frames on the input side, because the  
moment it starts clocking them out onto an overloaded link the goodput  
will plummet to zero.

ian