Re: random i/o speedup weirdness

On Wed, May 30 2007, Andras.Horvath_at_cern.ch wrote:
> > Instead I opted to try and fix the behaviour. It likely wont be perfect,
> > but some testing here shows that the IO at the end differs no more than
> > the noise during the entire run.
> >
> > So can I ask you to repeat the test with a current version of fio? I'd
> > suggest just grabbing:
> >
> > http://brick.kernel.dk/snaps/fio-git-latest.tar.gz
>
> Thank you very much, this produces a much flatter and more consistent
> result. The "summary screen" of fio is still wrong, but the "speedup"
> towards the end of the test is much, much less "extreme".

Good!

> root_at_pcitfiol25:/mnt/scratch/fio/new-fio# /usr/src/fio/fio sdc1.conf
> /dev/sdc1: (g=0): rw=randread, bs=4K-4K/4K-4K, ioengine=libaio, iodepth=2
> Starting 1 process
> Jobs: 1 (f=1): [r] [100.0% done] [ 0/ 0 kb/s] [eta 00m:00s]0s]
> /dev/sdc1: (groupid=0, jobs=1): err= 0: pid=15443
> read : io=7,534MiB, bw=3,081KiB/s, iops=752, runt=2563805msec
>
> [...]
>
> Run status group 0 (all jobs):
> READ: io=7,534MiB, aggrb=3,081KiB/s, minb=3,081KiB/s, maxb=3,081KiB/s, mint=2563805msec, maxt=2563805msec
>
> (why are min, max and average all the same values? or am I reading this
> wrong? the 'real' number would be about 700KiB/s for this disk)

This status is for all groups, as you only have one the min/max/etc
values will be the same. It's the same as the read value you have above,
3081KiB/sec. Is that not the correct number?

> Hmmm, on second thought I only just noticed that this is about the right
> value (from the cut-out part of the above output):
>
> bw (KiB/s) : min= 611, max= 1417, per=23.00%, avg=708.51,
> stdev=28.81

Hmm, please post the full output, thanks!

-- 
Jens Axboe