Re: [xfs] 610125ab1e: fsmark.app_overhead -71.2% improvement

[Dave Chinner]

On Mon, Sep 09, 2019 at 02:06:54PM +0800, Rong Chen wrote:
> Hi Dave,
> 
> On 9/9/19 1:32 PM, Dave Chinner wrote:
> > On Mon, Sep 09, 2019 at 09:58:49AM +0800, kernel test robot wrote:
> > > Greeting,
> > > 
> > > FYI, we noticed a -71.2% improvement of fsmark.app_overhead due to commit:
> > A negative improvement? That's somewhat ambiguous...
> 
> Sorry for causing the misunderstanding, it's a improvement not a regression.
> 
> 
> > 
> > > 0e822255f95db400 610125ab1e4b1b48dcffe74d9d8
> > > ---------------- ---------------------------
> > >           %stddev     %change         %stddev
> > >               \          |                \
> > >   1.095e+08           -71.2%   31557568        fsmark.app_overhead
> > >        6157           +95.5%      12034        fsmark.files_per_sec
> > So, the files/s rate doubled, and the amount of time spent in
> > userspace by the fsmark app dropped by 70%.
> > 
> > >      167.31           -47.3%      88.25        fsmark.time.elapsed_time
> > >      167.31           -47.3%      88.25        fsmark.time.elapsed_time.max
> > Wall time went down by 50%.
> > 
> > >       91.00            -8.8%      83.00        fsmark.time.percent_of_cpu_this_job_got
> > >      148.15           -53.2%      69.38        fsmark.time.system_time
> > As did system CPU.
> > 
> > IOWs, this change has changed create performance by a factor of 4 -
> > the file create is 2x faster for half the CPU spent.
> > 
> > I don't think this is a negative improvement - it's a large positive
> > improvement.  I suspect that you need to change the metric
> > classifications for this workload...
> To avoid misunderstanding, we'll use fsmark.files_per_sec instead of
> fsmark.app_overhead in the subject.

Well, the two are separate ways of measuring improvement. A change
in one without a change in the other is just as significant as
a change in both...

Cheers,

Dave.
-- 
Dave Chinner
dchinner@xxxxxxxxxx