Question

I just found System.currentTimeMillis is not accurate on windows XP, now I tried System.nanoTime() with the same code.

Since 1ms = 1,000,000ns, so I think the result should be 15,000,000ns, but it’s not.

See the sample code:

public class NanoTime {
    public static void main(String[] args) {
        long start = 0;
        long end = 0;
        while (true) {
            if (start == 0) {
                start = System.nanoTime();
            } else {
                long current = System.nanoTime();
                if (current != start) {
                    end = current;
                    break;
                }
            }
        }
        System.out.println("The time interval of your OS: " + (end - start) + "ns");
    }
}

The result is:

The time interval of your OS: 655ns

Seems it’s much better than System.currentTimeMillis(). But why? Can we believe this result?

Answer 1

If your Windows XP has SP3 then you can believe that result… on Windows the JVM implements nanoTime() internally using QueryPerformanceCounter/QueryPerformanceFrequency API which is basically located in HAL (Hardware abstraction layer) and uses some CPU-internal interval instructions which are rather accurate… it is only good for interval measurements NOT for time handling…

For a more elaborate description see http://blogs.oracle.com/dholmes/entry/inside_the_hotspot_vm_clocks and http://juliusdavies.ca/posix_clocks/clock_realtime_linux_faq.html

EDIT – as per comment:

The above mentioned function are used by the JVM itself internally when nanoTime is called which has absolutely NOTHING to do with JNI – they are not exposed to Java code!

EDIT 2 – as per comments:

nanoTime() is the right tool to measure elapsed time accurately while currentTimeMillis is the right tool to handle absolute time.