Why first call is so long?

Question

I found that first call of some function take too long time. Here my simple test:

public class MainTest {
    public static void main(String[] args) {
        long k = 0;
        for (int i = 0; i < 10; i++) {
            long start = System.nanoTime();
            k += doWork(i);
            System.out.println(System.nanoTime() - start);
        }
        System.out.println(k);    
    }

    public static long doWork(long var) {
        for (int i = 0; i < 100000; i++) {
            var += i;
        }
        return var;
    }
}

Results:

Can you give me some versions about why it's happening? Or some resources where I can find the answer. It can be connected with JIT, but I'am not sure about this.

I know that nanotime is not correct for testing. I try JMH:

# Warmup Iteration   1: 179791124,395 ops/s
# Warmup Iteration   2: 183962412,435 ops/s
# Warmup Iteration   3: 284320650,805 ops/s

The call to `nanoTime` probably takes longer than the code you are testing... — assylias, May 24 '17 at 09:22
In fact I just simplified code. I found this behavior when integrate groovy script in runtime and then just try the code above - same behavior. — learp, May 24 '17 at 09:25
Possible duplicate of [How do I write a correct micro-benchmark in Java?](https://stackoverflow.com/questions/504103/how-do-i-write-a-correct-micro-benchmark-in-java) — Turtle, May 24 '17 at 09:27
No, of course not. Why you think it's duplicate? I'am not trying to benchmark at all ! — learp, May 24 '17 at 09:41
It is close to a duplicate because your first problem is that you dont understand how to generate meaningful numbers. Beyond that: I appreciate the quick accept! — GhostCat, May 24 '17 at 09:55
When you say **meaningful** what you mean? :) I did not try **to generate meaningful numbers**. I interested why I get such results. — learp, May 24 '17 at 10:05
Well. You intend to draw conclusions from the numbers you generate. You ask "why do I observe xyz"; but what if your *observation* is not significant. But well, I think we all got something out of this. — GhostCat, May 24 '17 at 11:30
Yeah, I got your point. For these reason I used JMH(just in case). And test above is just simple example, just to explain the problem. — learp, May 24 '17 at 11:55

GhostCat · Accepted Answer · 2017-05-24T09:28:13.673

4

There are two aspects here:

Your benchmark is basically flawed and screwed up. You want to read this.
Yes, the JIT needs a warm up.

Simply spoken: at runtime, the JIT observes what your code is doing. It applies heuristics that trigger optimizations calls. Thus: don't expected deterministic behavior. It is hard to get to really meaningful numbers.

In order to get more reliable results, you will have to dramatically improve the quality of your "measurement" process.

edited May 24 '17 at 09:28

answered May 24 '17 at 09:26

GhostCat

137,827
25
176
248

Tbh, this question is a duplicate of your link, just like for the NPE questions. – Turtle May 24 '17 at 09:28
@GhostCat Thanks for the answer. Helped link to warm-up. It's not benchmark at all. I did not try to make **accurate** measurements. I get that behaviour and try to find out why this is so. – learp May 24 '17 at 09:58

Why first call is so long?

1 Answers1