El lun, 29-01-2007 a las 19:38 +0100, Petter Reinholdtsen escribió:
> [Ralf Gesellensetter] wrote:
> > on the German user list, there has been a thorough discussion about
> > performance optimization of NFS. Accordingly, udp (set by default)
> > connections are slow due repeated loss of data. Therefor it was
> > highly recommended to switch to TCP mode.
>
> I have problems understanding this. If there are a lot of IP packets
> lost, I fail to understand why it is better to leave the resending to
> the TCP stack instead of letting the NFS protocol take care of it.
> Can anyone explain it?
>
> > Maybe this is worth a wishlist entry?
> >
> > /etc/auto.master
> > => append ",tcp" to your option list
> >
> > otherwise, mount NFS shares with option "-o tcp".
> >
> > Did anybody else try this?
>
> We use TCP mounts to get past firewalls here at the university , but
> have not tried it to increase performance. I do know that it might
> help to increase the block size. Did you try this?
>
> Friendly,
> --
> Petter Reinholdtsen
>
Since two years ago we increase the block size, otherwise the perfomance
would be bad for schools where there are more than 300 PC mounting their
home directory on the nfs school.
This is our setup at clients:
echo "server_ip:/home /home nfs
rsize=16384,wsize=16384,timeo=20,retrans=10,intr,auto,rw 0 0"
>> /etc/fstab
We also add these lines:
echo 262144 > /proc/sys/net/core/rmem_default
echo 262144 > /proc/sys/net/core/rmem_max
at the start section of /etc/init.d/nfs-kernel-server
As I said before, it gives us good results, at least quite better than
the default options.
Regards.
José L.
Attachment:
signature.asc
Description: Esta parte del mensaje =?ISO-8859-1?Q?est=E1?= firmada digitalmente