brand new boxes running RH9 (downloaded a couple weeks ago, so it should be current) with integrated Intel GigE running on the PRO1000 drivers through a netgear 24-port unmanaged GigE switch (GigE on all ports). The boxes have single platter Hitachi 40GB HD's (8MB buffer, 7200 RPM) and 1GB of PC3200 RAM running dual channel. The FSB is running around 811 MHz with a 2.6 GHz Intel P4.

scp'ing a ~400MB file from one box to another takes about 15 seconds.
nfs reading the file takes about the same time
nfs writing the file takes about 12 minutes.

dd if=/dev/zero of=/mnt/home/testfile bs=16k count=256
takes just over 10 seconds (using the time command to judge)

I've tried changing the read & write block size from 8192 all the way to 32768 (I'm using nfs v2 & 3) - no change.

I've double checked the MTU settings (currently at 1500 on the card and via tracepath)

I've tried changing size of the socket input queue to 256K.

*To note: Every machine on my network is looking at 192.168.0.1 as a DNS server. 192.168.0.1 is not serving up DNS at the moment (new office - haven't setup a DNS server yet). Could this be causing problems? All machines in question have entries in eachother's hosts file.
I've disabled iptables. (didn't think that would make a difference)

Why is my nfs write going so slow?

edit:

/etc/exports on the server side is using the options: bg,noac,no_root_squash,no_subtree_checking,sync

I have five identical machines nfs mounted to one another (using different directories)

bump. anyone... anyone.... beuler... beuler?

I really don't know much about nfs - maybe I've overlooked something simple?

Well you seem to more about the network setup then I do but the one thing I would recommend changing is going from 'sync' to 'async'. From many of the sources I've read, you only need 'sync' if the data is crucial and must be preserved even if the system crashes. Since you're probably using a journaled filesystem (ext3) then data loss even for harddisk failure is fairly low anyhow.

Other then that, no idea. Very strange that you can get scp rw as fast and you get nfs ro.

I am faceing a similiar problem.

I am just running a 100MBit network but my reading perfomance is ok but if I want to write something on the NFS server I have around 60kb/s (no joke).
FTP works like a charm with 10MB/s though.

NFS and Samba are mainly using UDP right?
Could this be the problem? Actually UDP should be slighty faster then.
I am kinda desperate and don't know what to try anymore.
I have a realtek 8169 Gigabit card but it's connected to a 100MBit network and so running only as a 100MBit card.

Any hints or suggestions would be very welcome.

This link might be helpful.

http://www.acc.umu.se/~maswan/linux-netperf.txt