Re: Reopening jumbo frames in IS-IS

[Curtis Villamizar <curtis at faster-light.net>]

In message <845B4849-165E-4A2D-BBF7-508F84DC144A at cisco.com>
Fred Baker writes:


On Jul 15, 2005, at 10:05 AM, Radia Perlman wrote:
>  
> > What are the technical reasons that IEEE does not like large packets?
>  
> I can't speak for IEEE, but the reasons usually brought up include  
> implementation costs in terms of buffer depths, and mutual jitter  
> between competing traffic streams. If you have a session that sends a  
> packet every 20 ms and depends on that being mostly maintained,  
> having another session send packets that are 30 ms long and can get  
> several into the queue ahead of you can be a real pain. 


btw- 30 msec at 1 Gb/s is 30 mbit or about 4mB so the delay jitter
argument falls apart for GbE.  Its 400 kB for 100baseT.  It is a
problem for 10baseT and 40 kB in the queue.  For chips that put on the
order of 10 packets into an on chip hardware queue and don't consider
the length of the packets in the queue, 4-8KB MTU might get you 30
msec with 10baseT and 5-10 packets in the hardware queue.  The IEEE
ethernet types will remind us that a GbE and 10baseT might be bridged
together.  Still the jitter buffer argument is a very weak argument.

The memory requirements is not really an issue either.  You need the
same buffer depth (a lot more than a few packets) for any MTU.

Therefore the cost of changing the implementation is the only cost.
The cost of interoperability problems is an operational cost that
needs to be considered and is the reason to procede with caution.

Curtis