Why I am getting weird benchmark results between 0.99 and 100 percentile

Aleksey Shipilev shade at redhat.com
Thu May 6 08:51:15 UTC 2021

On 5/6/21 10:27 AM, Sarath Babu wrote:
> Weird means the score values are having huge differences between 0.9999 and 1.00 percentile,
> could you please help me to understand the reasons behind.
p1.00 is basically max, so it catches all the outliers. You probably have an outlier of 4.4+ 
seconds, which might be, for example, stop-the-world GC or something else.

Anyhow, even for normally distributed values it is "normal" to see higher percentiles to be orders 
of magnitude larger than the median/average values. Since the range of normal distribution is (-inf, 
+inf), it follows that with enough samples the p1.00 (which is max) for the normally distributed 
value is +inf. That is one of the reasons why "max" is rarely a good metric, and most people use the 
lower percentiles.


More information about the jmh-dev mailing list