Why standard deviation is always very high in percentages?

[MuhammadMunir12]

Can we limit the stdev value during test execution? Currently, the stdev values are like 86% which shows the variation in response time is very much greater. If anyone can tell, if we can limit its threshold value, please let me know.

[ddyurchenko]

Hi @MuhammadMunir12 ,

I was confused as much as you originally when I encountered that +/- StdDev is 60-90% in my benchmarks when Coefficient of Variation calculated manually is ~2%. After checking source code of wrk, I am no longer confused :)

+/- StdDev means what percentage of observations are within 1 standard deviation. For normal distribution, you would expect 68%. If it is larger than 68% - your distribution of values has tails larger than in normal distribution, if less than 68% - tails are smaller than expected (I stress "tails", because they can be on both side of distribution). Tbh it is not utterly useful metric, I suggest always running with --latency to see percentiles and know precisely how big is the right tail.