JMH - weird benchmarking results

Question

I'm using JMH to benchmark DOM parser. I got really weird results as the first iteration actually run faster than later iterations

Can anyone explain why this happens? Also, what do percentiles and all the figures mean and why it starts getting stable after the third iteration? Does one iteration mean one iteration of the entire benchmarking method? Below is the method I'm running

@Benchmark 
@BenchmarkMode(Mode.SingleShotTime) 
@OutputTimeUnit(TimeUnit.MILLISECONDS)
@Warmup(iterations = 13, time = 1, timeUnit = TimeUnit.MILLISECONDS)
public void testMethod_no_attr() {
    try {
        File fXmlFile = new File("500000-6.xml");
        DocumentBuilderFactory dbFactory = DocumentBuilderFactory.newInstance();
        DocumentBuilder dBuilder = dbFactory.newDocumentBuilder();
        Document doc = dBuilder.parse(fXmlFile);
    } catch (Exception  e) {
        e.printStackTrace();  
    }
}

Answer 1

These percentiles are horribly misleading. Don't use them.

If you have N = 10, it is no wonder all the bottom values are the same. Because you have way too few samples to talk about anything like a 99.999% quantile. The only value I would actually consider here is the median (50% quantile), and eventually the IQR (inter-quartile-ratio).

If you are optimistic, you'd assume your slowest sample is at 95%. If you are less optimistic, treat the slowest sample of N=10 as 90% quantile. And if you are more serious about these estimates, you may treat the spacing as being indicative of ca. 1/sqrt(N-1), i.e. assume that up to 33% of runs will be even slower than the slowest sample of N=10. You need many more samples to narrow down these estimates! Anything beyond 95% is just speculation. 99% - you cannot answer what is happening there. You just have not enough data.

As presented these values are just outright nonsense. Based on the data, you cannot estimate that P(99,9999)=something. This number corresponds to one out of a million runs being worse than this. But you only did 10, don't use this little N to predict what happens if you'd have a million. JMH should not print these extreme quantiles for small N.