[Beowulf] NFS over RDMA performance confusion

[Jonathan Barber]

On 13 September 2012 17:44,  <holway at th.physik.uni-frankfurt.de> wrote:
>
>> Ok.  Let me be clear, 10M as a file size is not useful as a test.  The
>> numbers are, for lack of any better way to describe this, meaningless.
>
> I am specifically testing the performance of NFSoRDMA vs NFSoTCP. I do not
> want my load to go to disk.
>
> What Im really curious about is why each process is limited to 7k or so IOPS.

I am not an iozone expert, but It could be that iozone is doing
synchronous IO (i.e. each write has to be acknowledged by the NFS
server). This would explain why the per-process IOPS is limited but
increasing the number of processes increases the total number of IOPS.

Try mounting the NFS export with the "async" option, and/or use the
async option for iozone to see if you get higher numbers.

Cheers
-- 
Jonathan Barber <jonathan.barber at gmail.com>