Re: Difference between copy to and copy from nfs mounted disk

rmacklem_at_uoguelph.ca
Date: 05/27/05 

Date: 26 May 2005 18:14:16 -0700

Zlatko Rek wrote:
> We have two identical Linux machines (Dual Xeon, Red Hat 8.0, custom
> kernel 2.6.9), connected with 1Gb Ethernet and with cross mounted disk
> partitions.
>
> I've found that time to copy 100MB file TO nfs mounted disk is appr.
> 200 times slower than copy FROM nfs mounted disk!
>
> For example:
> Copy over nfs from remote machine vortex3 to local machine vortex2
> (cp /home3/file .) takes 1.5 s, while copy from local machine vortex2
> to remote machine vortex3 (cp file /home3/) takes 305.8 s.
>
> Is this normal or is something wrong with our configuration?
>
> The following lines are from /etc/fstab:
>
> vortex2:
> /dev/hda3 /home2 ext3 defaults 0 0
> vortex3:/home3 /home3 nfs rsize=8192,wsize=8192,noac,nfsvers=3 0 0
>
> vortex3:
> /dev/hda3 /home3 ext3 defaults 0 0
> vortex2:/home2 /home2 nfs rsize=8192,wsize=8192,noac,nfsvers=3 0 0
>
> Any help is appreciated.
>
> Regards
>
> Zlatko
A couple of things you can try:
1 - do "netstat -s" on the server machine and look at the line
  "fragments dropped after timeout, in the Ip: section.
If those are happening, you need to reduce your wsize. (either
the server net interface or some intermediate router/switch is
dropping IP fragments when several are sent to the server, back
to back)
- alternately, try doing mounts over tcp. Those should work ok
   on Linux 2.6 kernels, I think?

You could also try a different net interface or look for a
different driver for the one you have or drop the net interface's
data rate down.

In general, I'd guess the client is transmitting too fast and
writes are getting dropped/retried. (You can also play with the
timeout value, but I recall the mount arg is only in the resolution
of seconds on a Linux mount.)

Good luck with it, rick
ps: Writing is usually slower, but not 200 X as slow.