Understanding bandwidth results

Understanding bandwidth results

[Andrew Davidoff]

Hi,

In fio output, I see two values for bandwidth that I'd expect to be
the same, but that are different. For example, in the results from a
sync randread fileio test:

read : io=182811KB, bw=18279KB/s, iops=18279, runt= 10001msec
...
    bw (KB  /s): min=  966, max=21410, per=99.59%, avg=18204.94, stdev=5773.57


I realize they are close, but why are the values for bw= and avg= not
the same? I want to make sure I understand if these are not actually
reporting on the same statistic.

Also, why is per= not 100%, if this job was run with only a single
thread in a single group?

Thanks.
Andy

Re: Understanding bandwidth results

[Jens Axboe]

On Sun, Jun 30 2013, Andrew Davidoff wrote:
> Hi,
> 
> In fio output, I see two values for bandwidth that I'd expect to be
> the same, but that are different. For example, in the results from a
> sync randread fileio test:
> 
> read : io=182811KB, bw=18279KB/s, iops=18279, runt= 10001msec
> ...
>     bw (KB  /s): min=  966, max=21410, per=99.59%, avg=18204.94, stdev=5773.57
> 
> 
> I realize they are close, but why are the values for bw= and avg= not
> the same? I want to make sure I understand if these are not actually
> reporting on the same statistic.
> 
> Also, why is per= not 100%, if this job was run with only a single
> thread in a single group?

That's just rounding and imprecise math. Should be fixed up though. It's
calculated in stat.c:show_ddir_status() and I'll be happy to take
patches :-)

-- 
Jens Axboe