IP Fragmentation

[KPS]

We seem to be having some unusual messages in the syslog on one of our hosts saying that the "NFS Server is not responding" and then it comes back and says that all is "ok."

We then put a sniffer on our VLAN that this server lives on and saw that there is
"IP Fragmentation out of Order" messages coming from the sniffer. Someone is recommending that we may want to increase the MTU Size on this host. We're wondering if this may be a good idea and/or if there are any better ideas on what to do next? If increasing the size of MTU is the way to go, can someone pass on how to do this task?

Thanks,
-Ken

[U.SivaKumar_2]

Hi,

Does this problem rise for the NFS clients in the local area network or WAN ?

If you are Increasing the MTU in the server, you have to increase the clients MTU to the same value .


regards,
U.SivaKumar

[Ron Kinner]

Changing the MTU might help a little but you are just masking the symptoms. I expect you could also get an "out of order" message if you drop packets so you might want to check
lanadmin
lan
display
(make sure you are on the interface you are interested in).

See if you are showing a lot of errors (bad cable or NIC or hub/switch) or a very high collision rate (mismatched duplex or babbling NIC). Also check netstat -s and see if you see a lot of errors or a high retransmission rate.

There are a ton of patches which have something to do with either NFS or ARPA. Make sure you have the latest recommended ones.

Ron

[rick jones · Accepted solution]

One can only increase the MTU if the NIC and all the infrastructure supports it. The HP GbE NIC support a larger MTU size, but the 100BT interfaces do not. Not all switches support large MTUs. Also, if you increase the MTU of one system, all other systems with which it will communicate at layer 2 (aka without going through a router) must be configured with the same MTU.

Out of order does indeed sound like fragments are being dropped, and the suggestion to check the lanadmin stats is spot-on - in particular, on the system(s) involved, you want to check for Outbound Discards and Outbound Errors.

You can also look at the receiving end of these IP fragments and see what the netstat -p ip statistics look like. If the fragments out of order are indeed for dropped fragments and not simply re-ordered fragments, then there should be increasing stats for fragment discard at the receivers.

If there is a lot of packet loss in the network, then switching to NFS over TCP might be the way to go - TCP will avoid fragmentation, and will generally handle lost segments (packets) better than NFS does when running over UDP.
It would be nice to know what sort of NICs are using

[KPS]

It looks like we're using the following type of Network Card:

HP PCI 10/100Base-TX Core

We have already updated to latest and greatest NFS and NIC Card patches to the recommendation of HP and are still seeing the errors...

[Jeff Schussele]

Hi Ken,

I'd start by trying another cable from the NIC to the switch. If still errors then try another port on the switch.
If yet still errors, I'd try another I/F on the system if you have an unused one available. The key is change ONLY one thing at a time to determine root cause.

Rgds,
Jeff

[rick jones]

When you say still seeing the errors, is that errors from the snifer, or have you run lanadmin on the sending side(s) and seen Outbound Discards and such?