Mixing jumbo and standard frames?

I'm looking at upgrading a Linux-based network to gigabit speed using 9K
jumbo ethernet frames. However, there are still some old 100Mbit computers
and devices that will still need to be connected to the network, these
needless to say do not support jumbo frame sizes.

It is my understanding that everything on the LAN needs to have the
same ethernet frame size. What is the best way to connect the jumbo and
non-jumbo devices? Web searching indicates that 802.1Q VLANs may
be the answer, but there is a lot of information to dig through and the
implementation details are not clear to me.

Any pointers on how to do this would be greatly appreciated - thanks!

--
Roger Blake
(Subtract 10s for email.)

Roger Blake <(E-Mail Removed)> wrote:
> I'm looking at upgrading a Linux-based network to gigabit speed
> using 9K jumbo ethernet frames. However, there are still some old
> 100Mbit computers and devices that will still need to be connected
> to the network, these needless to say do not support jumbo frame
> sizes.

> It is my understanding that everything on the LAN needs to have the
> same ethernet frame size. What is the best way to connect the jumbo
> and non-jumbo devices? Web searching indicates that 802.1Q VLANs
> may be the answer, but there is a lot of information to dig through
> and the implementation details are not clear to me.

If all your traffic is TCP, then you can "get away" with setting Jumbo
Frames on a subset of the systems. The reason is the TCP MSS (Maximum
Segment Size) exchange at the beginning of each connection. The
systems without JF enabled will specify a smaller MSS and all will be
well.

The issue arrises with UDP or direct link-level stuff, or anything
else without a TCP-like MSS exchange. Normal stuff like DNS will be
OK (IIRC) since that is already keeping itself below 1500 bytes
(handwaving). NFS over UDP however will have problems if you try to
go from a JF to non-JF system. The closest thing there to an MSS
exchange is the mount size and crippling to a 1024 byte mount size
would probably be worse than leaving the JF off.

Otherwise, the two systems have to be in separate IP subnets, with a
router in between them and/or you have to start adding a lot of host
routes with specific PathMTU's set on each JF-enabled system.

They don't necessarily have to be in separate VLANs. The "router" in
this case would have to be a JF-capable system, _perhaps_ with two
NICs but I think you can get away with logical interfaces instead - if
you can assign separate MTU's to each logical interface (eth0 and
eth0:1 in linux-like parlance). When JF talks to non-JF through the
router, it will fragment the IP datagrams to the proper size.

rick jones
--
No need to believe in either side, or any side. There is no cause.
There's only yourself. The belief is in your own precision. - Jobert
these opinions are mine, all mine; HP might not want them anyway...
feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...