conf t
system mtu 9198
wr
reload
Taken from cisco forums
The limitations of the platform in question would make no sense if
system mtu routing would apply to routed ports, but at the same time would not apply to SVIs. The point here seems to be that the model line
cannot fragment in hardware.
So to avoid that, it just makes sure every L3 interface (routed port
and SVI alike) has the same IP MTU, so there is never any need to
fragment a packet.
I've tested this in the lab using two 3560E chassis, trivially set up with the default VLAN, connected through Gigabit (so
system mtu jumbo
will apply) and then pinging each other's SVIs. Now let's configure
the following (using the maximum frame size the platform supports,
there's not much sense in limiting it here):
system mtu jumbo 9198
After rebooting, all physical ports running 1000Base or better will allow payloads of up to 9198 bytes to be encapsulated. The
routing mtu will still be 1500 at this point. Try to ping one switch from the other like
ping 10.1.1.2 size 1500 df-bit
and it will succeed, but increasing size to 1501 will fail.
Now let's add:
system mtu routing 9000
to both switches and try again (no reboots needed). You will see that
ping 10.1.1.2 size 6000 df-bit
will
suddenly work now, and the interface counters will make clear that no
fragmentation happens - it's really a single 6000 byte IP packet
bouncing forth and back between the switches. That works up to 9000,
and starts failing at 9001, exactly as you would expect.
Why is there a rumour that
system mtu routing doesn't apply to SVIs? Probably because
show interface of an SVI will show you an MTU of 1500 (or whatever your
system mtu
is), while the same command applied to a routed port will show 9000.
This seems to be a glitch, as so often with interface MTU in show
commands. More specifically, the
show interface MTU is
supposed to be the potential payload MTU of the underlying physical
interface of that routed interface, and there are other cases where it
displays rubbish. One should always compare to the
IP MTU as given by
show ip interface. Et voila: The IP MTU of our SVI (as given through
show ip int vl1) displays as 9000. So the succeeding ping is not a mystery and
system mtu routing does exactly what it intuitively states: Change the IP MTU of every L3 interface of the platform.
I
know this won't help in cases like the one discussed here, where
supposedly two L3 interfaces running at different MTUs are needed. In
such cases, one should first reassure that what's needed is really that
and there's no way to redesign the setup to avoid that (by placing L3
and L2 boundaries appropriately). If there is no way around that, the
3560 will likely have to go for something that has per-interface IP
MTU, like the 49xx or 4500X platforms.
Discussions about
MTU often mix up different problems and lead to chaos. IMO this is
because two things are often not regarded to the necessary extent by
the participants:
- There is not one MTU. When talking about
MTU, always define the layer you consider. That's often hard because
you actually have to think about a layer boundary, so two layers are
involved. The mythical 1500 for instance is the L3 MTU on top of a
classic L2 of the Ethernet family. Things fundamentally change when you
discuss L2 MTUs with regard to some underlying L1 (but things are
easier here, as L1+L2 are often developed together as one technology,
while the boundary between technology layers and the network layer has
more degees of freedom).
- MTU doesn't exist as such, but is an
emerging concept. In other words, MTU is what everybody in a system of
communicating nodes (typically a broadcast domain) thinks it is, with
the emphasis on everybody. In that sense it's like the
IP network that lives on top of a broadcast domain - it doesn't exist
per se, but by convention of everybody using that broadcast domain as a
bearer for IP packets sourced from non-colliding adresses from the
same visibility range (aka prefix, aka network and netmask, in ancient
times aka subnet/subnetmask). MTU is a convention as well, a single
node just changing its MTU (up or down from the convention, that is,
what everybody else uses) is a recipe for disaster. That's why you
don't do it except you know exactly what you are doing, and have the
might to change it everywhere in a broadcast domain (every involved
intermediate network device and every end node). It's viral. Luckily
it's not as widespread as the viral issue of people configuring ports
full duplex because that's what they always did, introducing duplex
mismatches up and down and then telling you that as obviously auto
negotiation doesn't work, they will continue this practice. You see the
common scheme: Breaking a convention is a bad idea, unless you are the
sole dictator.
