Accuracy of System.nanoTime() to measure time elapsed decreases after a call to Thread.sleep()

问题

I'm encountering a really unusual issue here. It seems that the calling of Thread.sleep(n), where n > 0 would cause the following System.nanoTime() calls to be less predictable.

The code below demonstrates the issue.

Running it on my computer (rMBP 15" 2015, OS X 10.11, jre 1.8.0_40-b26) outputs the following result:

Control: 48497
Random: 36719
Thread.sleep(0): 48044
Thread.sleep(1): 832271

On a Virtual Machine running Windows 8 (VMware Horizon, Windows 8.1, are 1.8.0_60-b27):

Control: 98974
Random: 61019
Thread.sleep(0): 115623
Thread.sleep(1): 282451

However, running it on an enterprise server (VMware, RHEL 6.7, jre 1.6.0_45-b06):

Control: 1385670
Random: 1202695
Thread.sleep(0): 1393994
Thread.sleep(1): 1413220

Which is surprisingly the result I expect.

Clearly the Thread.sleep(1) affects the computation of the below code. I have no idea why this happens. Does anyone have a clue?

Thanks!

public class Main {
    public static void main(String[] args) {
        int N = 1000;
        long timeElapsed = 0;
        long startTime, endTime = 0;

        for (int i = 0; i < N; i++) {
            startTime = System.nanoTime();
            //search runs here
            endTime = System.nanoTime();

            timeElapsed += endTime - startTime;
        }

        System.out.println("Control: " + timeElapsed);

        timeElapsed = 0;

        for (int i = 0; i < N; i++) {
            startTime = System.nanoTime();
            //search runs here
            endTime = System.nanoTime();

            timeElapsed += endTime - startTime;

            for (int j = 0; j < N; j++) {
                int k = (int) Math.pow(i, j);
            }
        }

        System.out.println("Random: " + timeElapsed);

        timeElapsed = 0;

        for (int i = 0; i < N; i++) {
            startTime = System.nanoTime();
            //search runs here
            endTime = System.nanoTime();

            timeElapsed += endTime - startTime;

            try {
                Thread.sleep(0);
            } catch (InterruptedException e) {
                break;
            }
        }

        System.out.println("Thread.sleep(0): " + timeElapsed);

        timeElapsed = 0;

        for (int i = 0; i < N; i++) {
            startTime = System.nanoTime();
            //search runs here
            endTime = System.nanoTime();

            timeElapsed += endTime - startTime;

            try {
                Thread.sleep(2);
            } catch (InterruptedException e) {
                break;
            }
        }

        System.out.println("Thread.sleep(1): " + timeElapsed);
    }
}

Basically I'm running a search within a while-loop which takes a break every iteration by calling Thread.sleep(). I want to exclude the sleep time from the overall time taken to run the search, so I'm using System.nanoTime() to record the start and finishing times. However, as you notice above, this doesn't work well.

Is there a way to remedy this?

Thanks for any input!

回答1:

This is a complex topic because the timers used by the JVM are highly CPU- and OS-dependent and also change with Java runtime versions. VMs may also limit the CPU capabilities they pass through to guests, which may alter the choices relative to a bare metal setup.

You may want to read the following, but note that some of those articles may be a bit dated, TSC reliability has improved a lot over the years:

OpenJDK Bug 8068730 exposing more precise system clocks in Java 9 through the Date and Time APIs introduced in java 8
http://shipilev.net/blog/2014/nanotrusting-nanotime/
https://btorpey.github.io/blog/2014/02/18/clock-sources-in-linux/
https://stas-blogspot.blogspot.de/2012/02/what-is-behind-systemnanotime.html
https://en.wikipedia.org/wiki/Time_Stamp_Counter (especially CPU capabilities, constant_tsc tsc_reliable nonstop_tsc in linux nomenclature)
there also are some platform-specific options that influence clock precision such as UseLinuxPosixThreadCPUClocks and AssumeMonotonicOSTimers
which timers get used also changes between JVM versions as newer hardware/OS capabilities become available
http://vanillajava.blogspot.de/2012/04/yield-sleep0-wait01-and-parknanos1.html

On linux you should check the tsc-related /proc/cpuinfo flags and the selected /sys/devices/system/clocksource/*/current_clocksource

回答2:

I can suggest at least two possible reasons of such behavior:

Power saving. When executing a busy loop, CPU runs at its maximum performance state. However, after Thread.sleep it is likely to fall into one of power-saving states, with frequency and voltage reduced. After than CPU won't return to its maximum performance immediately, this may take from several nanoseconds to microseconds.
Scheduling. After a thread is descheduled due to Thread.sleep, it will be scheduled for execution again after a timer event which might be related to the timer used for System.nanoTime.

In both cases you can't directly work around this - I mean Thread.sleep will also affect timings in your real application. But if the amount of useful work measured is large enough, the inaccuracy will be negligible.

回答3:

The inconsistencies probably arise not from Java, but from the different OSs and VMs "atomic-" or system- clocks themselves.

According to the official .nanoTime() documentation:

no guarantees are made except that the resolution is at least as good as that of currentTimeMillis()

source

...I can tell from personal knowledge that this is because in some OSs and VMs, the system itself doesn't support "atomic" clocks, which are necessary for higher resolutions. (I will post the link to source this information as soon as I find it again...It's been a long time.)

来源：https://stackoverflow.com/questions/33716359/accuracy-of-system-nanotime-to-measure-time-elapsed-decreases-after-a-call-to

标签

java

performance

nanotime