To Automount or to Not Automount?

To Automount or to Not Automount?

[Jon Forrest]

I run several Beowulf-style compute clusters
using CentOS 5. The question has come up of
whether having the compute nodes automount
needed filesystems is really necessary.
Using the automounter complicates other
aspects of configuring the cluster, but
the cluster software designers believe
that the benefits of automounting outweigh
the negatives.

In a Beowulf-style cluster there is a private
usually >= 1Gb/sec ethernet that handles the
NFS traffic. There are no routers involved.
If the network has problems that would cause
NFS to misbehave then the cluster as a whole
is probably in trouble. So, you can assume
a properly configured and functioning network.

So, I'd like to ask the NFS experts on this
list whether in a CentOS 5 environment with
NFS3 mounts over a functional private network,
would using static mounts result in any negatives
as compared to automounts. I'm especially interested
in overhead on the NFS server.

Cordially,

-- 
Jon Forrest
Research Computing Support
College of Chemistry
173 Tan Hall
University of California Berkeley
Berkeley, CA
94720-1460
510-643-1032
jlforrest-TVLZxgkOlNX2fBVCVOL8/A@public.gmane.org

Re: To Automount or to Not Automount?

[Thomas Haynes]

Jon Forrest wrote:
>
> So, I'd like to ask the NFS experts on this
> list whether in a CentOS 5 environment with
> NFS3 mounts over a functional private network,
> would using static mounts result in any negatives
> as compared to automounts. I'm especially interested
> in overhead on the NFS server.
>
> Cordially,

Is every node mounting every other node?

Or are you talking about every node mounting a central
server?

The static mounts would happen on boot and you might
have issues if every node was trying to mount every other
node. With automounts, you would delay the mounting
until the booting node has to access the other node.

I had this issue way back in 1998 with a Beowulf
cluster.  Heating issues forced me to shut down all
of the nodes at the same time and the interdependencies
in the mounts caused the bringup to be slow. :->

Re: To Automount or to Not Automount?

[Jon Forrest]

On 1/14/2010 12:58 PM, Thomas Haynes wrote:

> Is every node mounting every other node?

No. The compute nodes mount from the
one file server. The compute nodes
don't export anything.

> Or are you talking about every node mounting a central
> server?

Yes.

One person whom I respect told me that static mounts
result in more overhead on the server. I didn't
understand this but I don't claim to be
an NFS expert.


-- 
Jon Forrest
Research Computing Support
College of Chemistry
173 Tan Hall
University of California Berkeley
Berkeley, CA
94720-1460
510-643-1032
jlforrest-TVLZxgkOlNX2fBVCVOL8/A@public.gmane.org

Re: To Automount or to Not Automount?

[Trond Myklebust]

On Thu, 2010-01-14 at 13:15 -0800, Jon Forrest wrote: 
> One person whom I respect told me that static mounts
> result in more overhead on the server. I didn't
> understand this but I don't claim to be
> an NFS expert.

It depends upon your setup.

As Tom said, cluster bootup times can suffer if everybody is pounding
the server with static mount requests at the same time. However, once
that is done, there is usually very little overhead: if you have no
applications actually using the filesystem, then the client will
disconnect the TCP connection after ~5 minutes idle time (and on NFSv4,
it will stop renewing the NFSv4 leases). Once that is done, there is no
overhead whatsoever on the server.

However, one way in which static mounts can cause the server load to
increase is if you have applications whose behaviour is to follow active
mountpoints. For instance, the 'updatedb' daemon usually won't follow an
inactive automount point, but once the NFS filesystem is actually
mounted, it will traipse through, and index all the files it can find
there. Another application that often causes unnecessary traffic in the
static mount case is 'df'.

So before deciding on automount vs static mount, I'd advise you to do an
audit of your cluster nodes to see what applications are going to be
running and how their behaviour may cause the load to differ.

Cheers
  Trond