Quick performance question

[bogstad-e+AXbWqSrlAAvxtiuMwx3w at public.gmane.org (Bill Bogstad)]

On Fri, Aug 6, 2010 at 2:01 PM, Jerry Feldman <gaf-mNDKBlG2WHs at public.gmane.org> wrote:
> Currently I am using a RHEL 5.2 server to export a number of files via
> NFS. Simple. On the other servers, I use autofs to mount the
> directories. In preparation for moving the directories (about 800GB) to
> a dedicated NFS server, what I plan to do is:
> (1) rename the /mnts directory to /exports. This should have been done
> when I set up the server a few years ago. (The reason for the /mnts name
> is that I need to mirror Toronto.
>
> (2) then I want to mount the directories in /exports onto
> localalhost:/mnts. But, I want to use the same auto.mnts file on the nfs
> server that I do on the other 7 systems. So I can use the line:
> * ? ? ? ? ?-fstype=nfs,rw,nosuid,soft ? ? ?<nfs server name>:/exports/&
>
> I certainly can use
> * ? ? ? ? ?-fstype=nfs,rw,nosuid,soft ? ? ?localhost:/exports/&
>
> Is the Linux kernel smart enough to recognize that <nfs server name> is
> really the local host, and not send the bits out to the interface. I'm
> almost certain that the bits never go out on the network, but I just
> want to confirm it.

If <nfs server name> resolves to an IP address associated with some
real or virtual interface on the local machine, no packets will be
sent out on any physical interface.  (You might be able to set up some
very weird routing tables, to change this; but it won't happen with
any standard network configuration.)  This is true whether we are
talking about NFS, telnet, or any other network protocol.

OTOH, I think you might want to rethink doing it this way.  You'll
probably get better local IO performance on the server, if you don't
do force all local IO to those filesystems through NFS.   I haven't
kept up with benchmarks, but in the past Linux hasn't had good NFS
server performance.   I'm not sure what the best way to do this would
be, but you might look into manually setting up symbolic links or
possibly using the --bind option to the mount command.

Bill Bogstad