Synchronized java code performs times faster than unsynchronized one

Question

I work on a high concurrency app. In the app code I try to avoid synchronization where possible. Recently, when comparing test performance of a unsynchronized and synchronized code versions, it turned out synchronized code performed three-four times faster than its unsynchronized counterpart.

After some experiments I came to this test code:

private static final Random RND = new Random();
private static final int NUM_OF_THREADS = 3;
private static final int NUM_OF_ITR = 3;
private static final int MONKEY_WORKLOAD = 50000;

static final AtomicInteger lock = new AtomicInteger();

private static void syncLockTest(boolean sync) {
    System.out.println("syncLockTest, sync=" + sync);

    final AtomicLong jobsDone = new AtomicLong();
    final AtomicBoolean stop = new AtomicBoolean();

    for (int i = 0; i < NUM_OF_THREADS; i++) {
        Runnable runner;

        if (sync) {
            runner = new Runnable() {
                @Override
                public void run() {
                    while (!stop.get()){
                        jobsDone.incrementAndGet();

                        synchronized (lock) {
                            monkeyJob();
                        }

                        Thread.yield();
                    }
                }
            };
        } else {
            runner = new Runnable() {
                @Override
                public void run() {
                    while (!stop.get()){
                        jobsDone.incrementAndGet();

                        monkeyJob();

                        Thread.yield();
                    }
                }
            };
        }

        new Thread(runner).start();
    }

    long printTime = System.currentTimeMillis();

    for (int i = 0; i < NUM_OF_ITR;) {
        long now = System.currentTimeMillis();
        if (now - printTime > 10 * 1000) {
            printTime = now;
            System.out.println("Jobs done\t" + jobsDone);
            jobsDone.set(0);
            i++;
        }
        try {
            Thread.sleep(1000);
        } catch (InterruptedException e) {
            e.printStackTrace();
        }
    }

    stop.set(true);
}

private static double[] monkeyJob() {
    double[] res = new double[MONKEY_WORKLOAD];
    for (int i = 0; i < res.length; i++) {
        res[i] = RND.nextDouble();
        res[i] = 1./(1. + res[i]);
    }
    return res;
}

I played with the number of threads, workload, test iterations - each time synchronized code perfomed much faster than unsunchronized one.

Here are results for two different values of NUM_OF_THREADS

Number of threads:3
syncLockTest, sync=true
Jobs done 5951
Jobs done 5958
Jobs done 5878
syncLockTest, sync=false
Jobs done 1399
Jobs done 1397
Jobs done 1391

Number of threads:5
syncLockTest, sync=true
Jobs done 5895
Jobs done 6464
Jobs done 5886
syncLockTest, sync=false
Jobs done 1179
Jobs done 1260
Jobs done 1226

Test environment Windows 7 Professional Java Version 7.0

Here's a simillar case Synchronized code performs faster than unsynchronized one

Any ideas?

Answer 1

This is fascinating. I think @jtahlborn nailed it. If I move the Random and make it local to the thread, the times for the non-sync jump ~10x while the synchronized ones don't change:

Here are my times with a static Random RND:

syncLockTest, sync=true
Jobs done   8800
Jobs done   8839
Jobs done   8896
syncLockTest, sync=false
Jobs done   1401
Jobs done   1381
Jobs done   1423

Here are my times with a Random rnd local variable per thread:

syncLockTest, sync=true
Jobs done   8846
Jobs done   8861
Jobs done   8866
syncLockTest, sync=false
Jobs done   25956   << much faster
Jobs done   26065   << much faster
Jobs done   26021   << much faster

I also wondered if this was GC related but moving the double[] res to being a thread local did not help the speeds at all. Here's the code I used:

...
@Override
public void run() {
    // made this be a thread local but it affected the times only slightly
    double[] res = new double[MONKEY_WORKLOAD];
    // turned rnd into a local variable instead of static
    Random rnd = new Random();
    while (!stop.get()) {
        jobsDone.incrementAndGet();
        if (sync) {
            synchronized (lock) {
                monkeyJob(res, rnd);
            }
        } else {
            monkeyJob(res, rnd);
        }
    }
}
...

private static double[] monkeyJob(double[] res, Random rnd) {
    for (int i = 0; i < res.length; i++) {
        res[i] = rnd.nextDouble();
        res[i] = 1. / (1. + res[i]);
    }
    return res;
}

Answer 2

Random is a thread-safe class. you are most likely avoiding contention on calls into the Random class by synchronizing around the main job.