nfs write speed performance... still

[Antoine Jacoutot <ajacoutot_(_at_)_lphp_(_dot_)_org>]

Hi :)

I know some improvements on NFS were made with 3.6, but I'l still experiencing 
the same unsolvable problem I've been having for a while now.
I am no coder, so there's nothing much I can do to help resolve this issue, 
except testing, trying... anything. I would really like to use OpenBSD as an 
NFS server in production environment.

The problem I'm facing is that NFS writes are _really_ slow and, while copying 
a file to an NFS server, I cannot do anything anymore (since /home is also 
mounted on NFS). It looks like if my /home was mounted over a 56k connexion 
and that I was already downloading a file at full speed...

Here is my configuration ; note that I did not provide the dmesg since 
there're 2 machines involved, I did not want to make this message look to 
huge, but let me know if you need it.

** openbsd_nfs_server **
- OpenBSD-3.6
- xl0: flags=8843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> mtu 1500
        address: 00:04:75:f9:57:5e
        media: Ethernet autoselect (100baseTX full-duplex)
        status: active
        inet 192.168.1.1 netmask 0xffffff00 broadcast 192.168.1.255
        inet6 fe80::204:75ff:fef9:575e%xl0 prefixlen 64 scopeid 0x2
- nfsd_flags="-un 4"
- softdep activated on NFS exported filesystems

** openbsd_nfs_client **
- OpenBSD-3.6
- xl0: flags=8843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> mtu 1500
        address: 00:50:da:42:6e:8e
        media: Ethernet autoselect (100baseTX full-duplex)
        status: active
        inet 192.168.1.2 netmask 0xffffff00 broadcast 192.168.1.255
        inet6 fe80::250:daff:fe42:6e8e%xl0 prefixlen 64 scopeid 0x1
- mount_nfs options: rw,nodev,nosuid,-3,-U,-b,-i,-r8192,-w8192
- vfs.nfs.iothreads=4

Here is the test file that will be transfered.
$ du -ch file.test
76.4M   file.test

Now, first, let's copy the test file locally to eliminate hard drive problems.
--> local copy on openbsd_nfs_server
9.87s real     0.00s user     0.48s system # --> speed looks ok

Let's try scp transfer from openbsd_nfs_client to openbsd_nfs_server to 
eliminate network problems.
$ scp file.test openbsd_nfs_server:/tmp
file.test 100%   76MB   5.9MB/s   00:13 # --> speed looks ok

Now, the real test begins... NFS transfers.
--> read from openbsd_nfs_server to openbsd_nfs_client
15.07s real     0.00s user     0.98s system # --> speed looks ok
--> write from openbsd_nfs_client to openbsd_nfs_server
102.43s real     0.00s user     0.73s system # --> speed IS VERY SLOW

Now, I did some more testing with different OS involved. I won't post all the 
results, but here is the deal.
writes from:
linux --> openbsd => OK
openbsd --> linux => OK
freebsd --> openbsd => OK
openbsd --> freebsd => OK
openbsd --> openbsd => SLOW

So my problem is still NFS writing from openbsd to openbsd.
I tried forcing "100baseTX full-duplex" (which was autodetected) but no 
improvement was made. I'm not looking for huge transfer speeds, but one thing 
that does bother me is that, as my homedir in NFS mounted, I cannot do 
anything while writing to an NFS share: starting xterm takes about 30 
seconds.

I really need help on this and I'm ready to test anything you can propose.
Thanks for any hints.

Regards,

Antoine