Re: output - aggregate read bandwitdh

[Jens Axboe <axboe@kernel.dk>]

On Wed, May 22 2013, Edoardo Comar wrote:
> > > So 67bf982340d95ca98098ea050b54b4c7adb116c0 is the first bad commit,
> > > funky. Ah I think I see what it is, hang on.
> > 
> > http://git.kernel.dk/?
> > p=fio.git;a=commit;h=4b20c6814d4742e27a68f8a740a955283869fbcd
> > 
> > If you git pull, you'll get that. Please see if that makes it behave
> > like 2.0.13 for you.
> > 
> 
> Thanks,
> this has fixed the read aggrb reported value to be unreasonably high
> however the value is now significantly lower than 2.0.13  it's about 1/2
> while the write value is very comparable.
> Please find the outputs below.

This is due to the change in thinktime, fixing the unreasonable
latencies. Effectively it reduces your queue depth to what
thinktime_blocks is set to, which is 1 by default. Even if set to the
queue depth, you will still be completely draining the software queue
for every N thinktime_blocks. I'm guessing your device has write caching
enabled, hence the change is basically only apparent on reads. The
writes queue up nicely on the hw side still.

-- 
Jens Axboe