Is there something that will do this type of conversion?  Some formula
or rule of thumb?  Something other than just taking a time sample of
both sides and averaging it.

Thanks.

~F

[ Post a follow-up to this message ]

Faeandar wrote:
> Is there something that will do this type of conversion?  Some formula
> or rule of thumb?  Something other than just taking a time sample of
> both sides and averaging it.

Please define what an nfs op is and what an iop is.

[ Post a follow-up to this message ]

Mike Eisler wrote:
> Faeandar wrote: 
>
> Please define what an nfs op is and what an iop is.

'Way back when, nhfsstone was the way one benchmarked the performance of
an NFS server, but because of some deficiencies (such as dependencies on
client specifics and general lack of rigor in definition) a benchmark
called LADDIS (an acronym created from the names of the companies
involved in creating it:  Legato, Auspex, DG, DEC, Interphase, and Sun)
superseded it and (IIRC) introduced the concept of 'nfs ops' (each of
which often required varying numbers of disk accesses to complete and/or
might benefit from cached data).  LADDIS was incorporated into the SPEC
SFS server benchmark, which may still be the standard nfs measurement today.

IOPS, of course, refers to simple disk accesses (i/o operations per
second), either of a single disk or of an aggregation of them - or to
'effective' disk accesses when applied to a server which may satisfy
some out of an internal cache.

LADDIS does define the mix of nfs operations to perform, and one could
presumably derive the number of disk accesses they would require in the
absence of caching - but even then that number would still depend upon
the specific server file system implementation, whereas any IOPS
measurement should be essentially independent of anything but the
hardware.  Therefore, it is not clear that any stable relationship
exists between nfs ops and iops (at least if Faeandar was referring to
LADDIS-style nfs ops).

- bill

[ Post a follow-up to this message ]

On 19 Feb 2006 09:57:53 -0800, "Mike Eisler" <spamisevi1@yahoo.com>
wrote:

>
>Faeandar wrote: 
>
>Please define what an nfs op is and what an iop is.

Bill is going in the right direction.  IOPS would be the OS operations
against locallly owned disks (or controllers or Lun's or whatever)
whereas NFS ops are operations made against the NFS stack/client.

I thought the difference was generally understood but maybe not.  What
was the reason for your question?  Is there theoretically no
difference between them?

~F

[ Post a follow-up to this message ]

Faeandar wrote:
> On 19 Feb 2006 09:57:53 -0800, "Mike Eisler" <spamisevi1@yahoo.com>
> wrote:
> 
>
> Bill is going in the right direction.  IOPS would be the OS operations
> against locallly owned disks (or controllers or Lun's or whatever)
> whereas NFS ops are operations made against the NFS stack/client.

What lengths are passed to the OS operations?

> I thought the difference was generally understood but maybe not.  What
> was the reason for your question?  Is there theoretically no

I don't know what you think an IOop or NFSop is.

> difference between them?

If Bill is talking about "LADDIS" (really, SPEC SFS), then you can
look at the documentation to see which NFS operations in the workload
are reads and writes, assuming that's what you meant. Just take the
reported SPEC SFS ops/sec and ultiply by the percentage of ops that
are the sum  of read and write, and you get your number.

http://www.spec.org/sfs97r1/docs/chapter1.html#table1

I don't see in the documents the read and write transfer size. I think
it is 8 Kbytes. Assuming so, if your IOP uses 8KByte transfers, then
you can directly translate from SPEC to NFS.

However if you simpy want to see what the maximum IOPs are for a given
NFS server, then you'd need to run the benchmark with 50% reads and
50% writes.

So again, you need to define your terms.

[ Post a follow-up to this message ]

Mike Eisler wrote:

> http://www.spec.org/sfs97r1/docs/chapter1.html#table1
>
> I don't see in the documents the read and write transfer size. I think
> it is 8 Kbytes. Assuming so, if your IOP uses 8KByte transfers, then
> you can directly translate from SPEC to NFS.

It is 8KB. Some day SPEC will make it 32KB, or better, dynamic. Although
that won't matter too much as most accesses are <= 16KB (~ 85%, with
almost half <= 8KB).

In the real world, depending on what clients and servers you
run, the transfer size of a large sequential I/O may
vary from 8KB to 1MB.

So to re-ask Mike's question, what is your NFSop?

-David

[ Post a follow-up to this message ]

On Sun, 19 Feb 2006 23:15:56 GMT, Faeandar <mr_castalot@yahoo.com>
wrote:

>On 19 Feb 2006 09:57:53 -0800, "Mike Eisler" <spamisevi1@yahoo.com>
>wrote:
> 
>
>Bill is going in the right direction.  IOPS would be the OS operations
>against locallly owned disks (or controllers or Lun's or whatever)
>whereas NFS ops are operations made against the NFS stack/client.
>
>I thought the difference was generally understood but maybe not.  What
>was the reason for your question?  Is there theoretically no
>difference between them?
>
>~F

Ok, all good questions and things I had not considered (blinders on
right now).

In the case of a DAS setup it's simple enough to measure IO
performance with Iozone or Iometer or whatever.  With NAS it's simple
enough to measure IO performance with Iozone or SpecFS.  But if I want
to measure IO performance from end to end (client NFS requests to
NAS's requests of it's own disks) what are some recommendations for
doing that?

My thought was that I could run some benchmarks with an NFS tool and
capture those metrics, then run the same test again but this time from
the NAS server.  An example:

Iozone from an NFS client with specific file sizes, record sizes,
read/write variabls, etc.  Then the same spec's on the NAS server to
it's local disk.

My concern was the correlation of the two.  Maybe I'm over
complicating this, but suggestions are welcome.

Thanks.

~F

[ Post a follow-up to this message ]

"Faeandar" <mr_castalot@yahoo.com> wrote in message
 news:i53qv1hjs25nfq5a084ke1att9o8qc0rqc@
4ax.com...
> On Sun, 19 Feb 2006 23:15:56 GMT, Faeandar <mr_castalot@yahoo.com>
> wrote:
> 
>
> Ok, all good questions and things I had not considered (blinders on
> right now).
>
> In the case of a DAS setup it's simple enough to measure IO
> performance with Iozone or Iometer or whatever.  With NAS it's simple
> enough to measure IO performance with Iozone or SpecFS.  But if I want
> to measure IO performance from end to end (client NFS requests to
> NAS's requests of it's own disks) what are some recommendations for
> doing that?
>
> My thought was that I could run some benchmarks with an NFS tool and
> capture those metrics, then run the same test again but this time from
> the NAS server.  An example:
>
> Iozone from an NFS client with specific file sizes, record sizes,
> read/write variabls, etc.  Then the same spec's on the NAS server to
> it's local disk.
>
> My concern was the correlation of the two.  Maybe I'm over
> complicating this, but suggestions are welcome.
>
> Thanks.
>
> ~F

Perhaps one could find a correlation between Iozone running
on an NFS client (capture NFS traffic with a tcp dump )
and Iozone running on the local disks in the NAS ?
This will not resolve the issue of defining the terms NFSop
and Iop, but it might give the correlation that you are
looking for :-)

Enjoy,
Postmaster

[ Post a follow-up to this message ]