Sample time with "inverted" histogram?

[Aleksey Shipilev]

Hi Chris,

On 11/27/2015 07:08 PM, Chris Vest wrote:
> I sometimes find myself wishing the sample time benchmark mode had an “inverse” histogram, because I find that this makes it easier to spot when response times are modal.
> 
> To demonstrate what I mean, the current histogram output looks like this: (numbers are fictive)
> Samples, N = 3184
>       mean =      2.531 ±(99.9%) 0.014 us/op
>        min =      0.119 us/op
> p( 0.0000) =      0.119 us/op
> p(50.0000) =      2.420 us/op
> p(90.0000) =      2.644 us/op
> p(95.0000) =      2.816 us/op
> p(99.0000) =      4.408 us/op
> p(99.9000) =     13.152 us/op
> p(99.9900) =     36.901 us/op
> p(99.9990) =    254.240 us/op
> p(99.9999) =    260.864 us/op
>        max =    260.864 us/op
> 
> And I would like a configuration option that would make it look more like this: (numbers even more fictive)
> Samples, N = 3184
>       mean =      2.531 ±(99.9%) 0.014 us/op
>        min =      0.119 us/op
>      us/op = samples added in increment
>          1 = 1081
>          2 = 194
>          4 = 959
>          8 = 357
>         16 = 3
>         32 = 3
>         64 = 118
>        128 = 410
>        256 = 58
>        512 = 1

> This way, I can more easily see that the numbers in my made-up
> example fall into 2 or 3 groups. When I have needed something like
> this, I have thus far reached for HdrHistogram and its
> LogarithmicIterator, and the
> HistogramIterationValue.getCountAddedInThisIterationStep method.

It might be worthwhile to explore. JMH's SampleBuffer is recording the
HDR data, so those histogram values are available. Therefore it is a
question of presentation: maybe print the histogram alongside with
quantiles?

Thanks,
-Aleksey