Switch/Server MTU Jumbo Frames

[aheckman]

I have a 3560G switch that all of our servers are plugged into. Most of the servers are HP DL380-G3 or G5's. Because of that i have the HP Network Config tool which allows me to set the MTU as well as do some nice LCAP etherchannel. I have the etherchannel channel working nicely on some test servers and i would like to try out the higher MTU and see performance gains if any. However i know that i need to set the system mtu, and cannot change per interface.

My question is that if I change the system mtu jumbo to 9000 will that break the servers that i haven't changed to the larger MTU? Likewise what about my switch to switch point-to-point links where i haven't changed the MTU on that switch yet.... I would rather not break anything, but at the same time want to only test with a small number of devices not everything at once.

[Giuseppe Larosa]

Hello Andrew,

the switch doesn't generate traffic on its own.

you can have performance gains on communications between pairs of servers that support the extended MTU

TCP session setup contain the MSS parameter so two servers with mismatching MTU should be able to communicate using TCP sessions

Hope to help

Giuseppe

[aheckman]

So to diagram your thought out...

Server with high MTU sends session startup, and hits switch with high MTU, so no issues there, then Switch forwards traffic to 1 of 2 places:

1.) another server with high MTU set = OK no issues...

2.) Server withOUT high MTU set = possible performance issues?

What about Server sending traffic to clients downstream that go

Server (High) > Switch (High) > Switch (Low) > Client (Low)

[Joseph W. Doherty]

In theory nothing should "break". In practice it can when path MTU isn't properly handled. Also what can happen, if hosts use a larger MTU than the network infrastructure supports, performance can be degraded vs. using the smaller MTU supported by the infrastructure to start with.