NFS changes between 2.6.8 and 2.6.24

We are looking into some unusual NFS behavior changes as we migrate from a
2.6.8 kernel to a 2.6.24 kernel.

We have two server farms, each having a few dozen machines accessing a
central NAS via NFS. For testing purposes we have a 2.6.8 machine and a
2.6.24 machine both running webmail, therefore making frequent NFS access
to small files. When we look at the traffic from the 2.6.8 kernel we see
what we consider to be a reasonably low amount of NFS traffic. The traffic
also follows what appears to be a reasonable pattern: LOOKUP calls, ACCESS
calls, READDIRPLUS calls, an occasional GETATTR call. I'd estimate that
GETATTR calls make up 10-15% of the total NFS traffic.

The same webmail app running on the 2.6.24 kernel generates a lot more NFS
traffic, and it's not nearly as intuitive a pattern: dozens of GETATTR
calls, on occasional LOOKUP, ACCESS, REMOVE, etc. I'd estimate that the
GETATTR calls account for easily 90% of the total NFS traffic. Machines
running this new kernel are placing a much higher load on the NAS and the
internal network the NAS runs on than the machines running the older
kernel do.

We've reviewed the kernel change logs and noted a few comments on minor
changes in the NFS code, but we haven't seen any comments that seem to
explain this kind of change in NFS performance. Can anyone point me to a
source of info on this? We'd love to migrate all the farm machines to the
newer kernel, but until we get a handle on the change in NFS behavior we
can't really move forward.

Daniel Mahoney wrote:
> We are looking into some unusual NFS behavior changes as we migrate from a
> 2.6.8 kernel to a 2.6.24 kernel.
>
> We have two server farms, each having a few dozen machines accessing a
> central NAS via NFS. For testing purposes we have a 2.6.8 machine and a
> 2.6.24 machine both running webmail, therefore making frequent NFS access
> to small files. When we look at the traffic from the 2.6.8 kernel we see
> what we consider to be a reasonably low amount of NFS traffic. The traffic
> also follows what appears to be a reasonable pattern: LOOKUP calls, ACCESS
> calls, READDIRPLUS calls, an occasional GETATTR call. I'd estimate that
> GETATTR calls make up 10-15% of the total NFS traffic.
>
> The same webmail app running on the 2.6.24 kernel generates a lot more NFS
> traffic, and it's not nearly as intuitive a pattern: dozens of GETATTR
> calls, on occasional LOOKUP, ACCESS, REMOVE, etc. I'd estimate that the
> GETATTR calls account for easily 90% of the total NFS traffic. Machines
> running this new kernel are placing a much higher load on the NAS and the
> internal network the NAS runs on than the machines running the older
> kernel do.
>
> We've reviewed the kernel change logs and noted a few comments on minor
> changes in the NFS code, but we haven't seen any comments that seem to
> explain this kind of change in NFS performance. Can anyone point me to a
> source of info on this? We'd love to migrate all the farm machines to the
> newer kernel, but until we get a handle on the change in NFS behavior we
> can't really move forward.

I don't know what has changed in NFS between 2.6.8 and 2.6.24, but the folks
at linux-nfs@vger.kernel.org would. That is the mailing list of the NFS
developers. You should post your message there.

Larry

On Fri, 16 May 2008 21:02:54 +0000, Larry Finger wrote:

> Daniel Mahoney wrote:
>> [quoted text muted]
>
> I don't know what has changed in NFS between 2.6.8 and 2.6.24, but the folks
> at linux-nfs@vger.kernel.org would. That is the mailing list of the NFS
> developers. You should post your message there.
>
> Larry

Thanks, I'll do that.