how to enable path MTU discovery for UPD-packets (HP-UX 10.20)

[MARREEL Chris_1]

Hello,

I have a rather old HP-UX station (B132L) with HP-UX 10.20. I cannot upgrade to a newer version of HP-UX because the software is only supported on HP-UX 10.20.
This workstation must communicate with another station using NFS over a WAN-connection (VPN over internet).

This was working, until the ISP changed a CORE Router, and the MTU (Maximum Transfer Unit) was reduced, and over my NFS-mount cannot copy large files (larger than 1300bytes).
So a small file (smaller than +- 1380bytes) can be copy-ed (content of the file is transferred in one small frame), but larger files cannot be copy-ed (because a larger file is split in multiple frames of 1514 bytes, and these cannot be transported over this WAN-connection).

I have discovered (in the MAN-page of route on the HP-UX 10.20 machine) that I can specify a PATH MTU (route -p pmtu add host ....), but in that same MAN-page there is a paragraph :

By default, Path MTU Discovery is enabled for TCP sockets and disabled for UDP sockets.

My question : how can I ENABLE the MTU Discovery for UDP sockets (because NFS uses UDP Sockets) ?

Thanks,
Chris MARREEL

[Steven E. Protter]

Shalom Chris,

You should be able to set this with the ndd util and make it permanent at boot with the /etc/rc.config.d/nddconf file.

The help for ndd should let you find the parameter to set.

SEP

[MARREEL Chris_1]

Hello Steven,

where can I find more info on 'ndd', because on the HP-UX 10.20 box "man ndd" is not found...

Can you point me in the right direction ?

Thanks,
Chris MARREEL

[Matti_Kurkela]

I could not find a way to enable the Path MTU discovery for UDP on HP-UX 10.20, but as a work-around, you might use the "-M" option of the lanadmin command to set the MTU of the interface to a value that's acceptable to your ISP.

Lowering the MTU of your network interface will effect all traffic going through the interface, not just the part of the traffic that goes through the ISP. If the ISP's new MTU value is not much smaller than the standard value of 1500, the loss of performance is probably not significant.


By the way, if the Path MTU discovery is not used, the network devices along the path should fragment the packets if they're too big for some part of the path.

So the packets should still go through, unless the NFS sets the "Don't Fragment" bit on the packets (as fragmentation will hurt performance), or something is blocking the fragmented packets. Perhaps there is an overzealous firewall rule somewhere?

[Bill Hassell · Accepted solution]

The ndd command is called nettune in your old OS version. nettune with no values may show you the parameters, or you can look at the man page for nettune. Once you set nettune, you'll need to change the config file in /etc/rc.config.d so the new parameter will be reapplied when you bootup.

[MARREEL Chris_1]

Guy's thanks for your quick answers.
I can indeed use nettune to manipulate udp_pmtu and tcp_pmtu.

Thanks,
Chris MARREEL

[rick jones]

While you may be able to change the settings for UDP pamth MTU as suggested via nettune, when your ISP changed their core router, as the DF bit was not set on the IP datagrams carrying the UDP datagrams carrying the NFS messages, it should have fragmented them to the required size and forwarded them. They then should have been reassembled at the destination.

If the router is not fragmenting, then it is out of spec and needs to be fixed.

If some of the datagrams are being lost, then you may have a basic packet loss problem.

While by default, Path MTU discovery will not be done on UDP datagrams, I _thought_ that when one initiated communications, that the PathMTU discovery information in the PATH MTU route would be honored even by UDP. It was just that IP datagrams carrying UDP datagrams would not have the DF bit set.

Also, as this is NFS over a WAN you may want to consider setting your mount size to something less than the MTU of the path and avoid the fragmentation altogether.

FWIW, while different apps vendors may not "support" their 10.20 binaries on 11.X, euros to berliners (aka dollars to doughnuts) those apps will likely "work" on 11.X (unless they have hard kernel dependencies)