[c-nsp] Suspect MTU Issues

[Phil Mayers]

On 07/13/2011 10:13 PM, Leigh Harrison wrote:
> This discussion brings me neatly onto my follow on question then:-
>
> On the ME3600X switches they will allow me to set interface mtu of up to 9800 bytes. Some of my team are arguing that we only need 1548, some are saying 1600.
>
> We've got dark fibre, so should we be going for the maximum mtu size possible on the box (taking into account max mtu of the box it plugs into), or is there a good "all rounder" for mtu size?
>
> What mtu will cause me the least pain in years to come?

In an ideal world, you pick the largest value that all your kit can 
support (e.g. we use 9212 as our untagged ethernet MTU, 9100 as our IP 
MTU). Too big is not harmful.

However, you often find some links can't support that big an MTU, and 
they end up running e.g. 1530. Provided that path MTU discovery works 
(so that multihop protocols like iBGP and LDP don't get blackholed) this 
works OK in my experience, but I've heard of others having problems, so 
you may feel safer with a consistent, lower value.

The important thing is that your network MTU exceeds your payload MTU by 
at least the transport overhead. Bearing in mind the large diversity of 
payload MTUs, depending on what services you're running.

If for example you are doing EoMPLS of normal 1500-byte ethernet with no 
or one vlan tag, the payload can be up to 1518 bytes, and the overhead 
is 2 MPLS labels, giving 1524.

OTOH L3VPNs of 1500-byte IP traffic typically need 1508.

But in principle other MPLS services like TE can add one or more labels, 
and of course you may find you need to move jumbos as payload at a later 
date...