Problem in using System.currentTimeMillis(); in Java

Question

I tried to observe the time taken by different inputs in the calculation of the nth Fibonacci number, but the output is <50ms for the first input and 0 ms for the rest Why so?

import java.io.*;
import java.util.*;
class fib{
    long fibo(int s){
        if(s==1 ||s==2)
        return 1;
        else return fibo(s-1)+(s-2);
    }
}
class fibrec{
    public static void main(String args[]) throws java.io.IOException{
        BufferedWriter wr=new BufferedWriter(new FileWriter("C:/Users/91887/desktop/books/java/foo3.txt"));
        fib f=new fib();
        Random rand=new Random();
        int input[]=new int[10];
        for(int p=0;p<10;p++){
            long st=System.currentTimeMillis();
            int i=rand.nextInt(12000);
            wr.write("Input : "+i+"\nOutput : "+f.fibo(i)+"\n");
            long et=System.currentTimeMillis();
            wr.write("Time taken = "+(et-st)+"ms\n\n");
            System.out.println(st+"\t"+et+"\t"+(et-st));
        }

        wr.close();
    }
}

Answer 1

The granularity of the millisecond clock is at best one millisecond ¹ .

But apparently, the execution times for your loop iterations are less than one millisecond. Sub-millisecond time intervals cannot be measured accurately using System.currentTimeMillis() . That is why you are getting zeros.

The explanation for the first measurement being 35 milliseconds is that this is due to JVM warmup effects. These may include:

• time taken to load and initialize library code ² ,
• time taken to JIT compile code, and
• time taken up with a (possible) GC during or after loading and JIT compilation.

Secondly, I notice that your time measurement includes the time taken to print the number. You should move that after the second call to get the clock value because it could be significant.

Finally, it you want to get reproducible results, you should explicitly seed Random yourself rather than relying on the OS to give you a random seed. And I'm not convinced that you should be benchmarking a Fibonacci algorithm with random inputs anyway ...

---

^{1 - The numbers in your output suggest that it is actually 1 millisecond ...}

^{2 - For example, the initialization and construction of a Random instance entails an OS call to get some "entropy" to seed the random number generator. That should be fast, but there are circumstances where it may not be. In your case, this happens before you start measuring time ...}

Answer 2

The code between the two calls to System.currentTimeMillis() is executing too fast (after the first iteration) for any difference to be captured. You'd be able to see a difference if you were using System.nanoTime() .

As for why the first iteration is slower than the subsequent ones, that would be because Java uses a Just In Time (JIT) compiler to optimise code at runtime.