How many runs of Java program do we need to warm-up the JVM?

Question

Suppose I have a Java program Test.class. I want to measure its execution time. I wrote a wrapper to do this as below:

class RunTest {

    public static void main(String[] args) {

        long sum = 0;
        int iterations = 20;
        int warmupNum = 10;

        for(int i=0; i<iterations; i++){

            long start = System.nanoTime();
            Test.main(args);
            long end = System.nanoTime();

            if( i > warmupNum )
              sum += end - start;
        }

       System.out.println("ave: "+sum/(iterations-warmupNum));
    }
}

Here how to choose warmupNum, the larger the better? How large is enough? Is this a "standard/common" way to measure Java program's performance?

Countdown

10,000 is the default value before the JIT starts doing its job. You might be interested in reading this post. – Subkingdom

@Subkingdom You are saying the 10000 is the number of repetitions for optimizing a method, right? Here, I want to know of warmup iterations of program to get a stable execution time. – Countdown

yes - the JIT will start compiling a method after it's been called 10,000 times. You can change that value with the parameters you pass to the JVM but I can't find it right now. – Subkingdom

this gives more info - 10000 is in server mode - it seems to be 1500 in client mode. – Subkingdom

@Countdown Accept my answer if I've answered your question. – Selfinductance

Answer 1

It is better to use Caliper than creating your own micro-benchmark utility.

How do I write a correct micro-benchmark in Java?

Selfinductance