NFS problems with the SuSE Linux Enterprise Server 8 for i- and pSeries
Support knowledgebase (nfs_udp_sles8ppc)
Applies to
SuSE Linux PowerPC: Version 8.1
Symptom:
In some rare scenarios there is a small possibility of a data
integrity problem when using NFS over UDP with standard
parameters. The NFS server has to have high bandwidth links of more
than 200 MBit/sec total for outbound data. In addition this bandwith
has to be in use for the window of possibility to open. Services that
run over TCP are not affected.
Cause:
There is a general problem with data fragmentation/reassembly within
the IP definition. Fragmentation timeout and transport ratio has to be
in a certain relation to be 100% safe. Please notice that all
implementations of NFS over UDP are hit by that problem.
Solution:
-
Set down IP fragmentation timeout to three seconds:
echo 3 > /proc/net/sys/ip_frag_timeout
This might cause problems when connecting to an other site
over a slow latency WAN.
-
Avoid fragmentation: If you're running NFS over a line that
supports jumbo frames please use them. Most GBit adapter and some
100MBit adapter should already do. To set that on e.g. eth4, type:
ifconfig eth4 mtu 9000
-
If both of the above settings are not fit your environments
you will have to reduce read/write block sizes to 1KBytes for NFS
mounts.
Add the following options to your /etc/fstab or were mount options
are defined (NIS/LDAP depending on your setup):
rsize=1024,wsize=1024
Please be aware that there you will find an update OS kernel in
the near future. This will give you the possibility to adjust
ip_frag_timeout per outgoing device besides a couple of other changes
that even lower the possibility of data corruption.
Keywords: NFS, UDP, PPC, ISERIES, PSERIES
SDB-nfs_udp_sles8ppc, Copyright SuSE Linux AG, Nürnberg, Germany
- Version: 20. Mar 2003
SuSE Linux AG - Last generated: 20. Mar 2003 by snbarth (sdb_gen 1.40.0)