inefficient threads in java

Question

I currently have some problems to understand why in some cases, parallelization in Java seems infficient. In the following code, I build 4 identical tasks that are executed using a ThreadPool.

On my Core i5 (2 core, 4 thread), if I set the number of workers to 1, the computer needs around 5700ms and use 25% of the processor. If I set the number of workers to 4, then I observe 100% of CPU usage but... the time of computation is the same: 5700ms, while I expect it to be 4 times lower.

Why? Is it normal?

(Of course my real task is more complicated, but the example seems to reproduce the problem). Thank you by advance for your answers.

Here is the code:

public class Test {

public static void main(String[] args) {
    int nb_workers=1;
    ExecutorService executor=Executors.newFixedThreadPool(nb_workers);
    long tic=System.currentTimeMillis();
    for(int i=0; i<4;i++){
        WorkerTest wt=new WorkerTest();
        executor.execute(wt);
    }
    executor.shutdown();
    try {
        executor.awaitTermination(1000, TimeUnit.SECONDS);
    } catch (InterruptedException e) {e.printStackTrace();}
    System.out.println(System.currentTimeMillis()-tic);
}

public static class WorkerTest implements Runnable {
    @Override
    public void run()  {
        double[] array=new double[10000000];
        for (int i=0;i<array.length;i++){
            array[i]=Math.tanh(Math.random());
        }
    }
}
}

Just to verify, when you say "set the number of workers" are you referring to the number of threads in the executor service, the number of tasks submitted to the executor service, or both? — Alex, May 14 '14 at 19:27
I think he's obviously talking about the hard coded `4` in the for loop. — pamphlet, May 14 '14 at 19:29
I am referring to the number of threads in the executor service. The number of tasks submitted stays the same (4). — SamuelHumeau, May 14 '14 at 19:30
I think it's equally likely he could be referring to the `nb_workers` variable. — Alex, May 14 '14 at 19:30
I can't reproduce this. Setting `nb_workers` to 1 will take ~7s. Setting it to 4 takes ~2s. — Thomas Jungblut, May 14 '14 at 19:31
@EWit, the question is ambiguous in this regard. OP has clarified that the number of tasks varies, not the number of threads. — pamphlet, May 14 '14 at 19:39

score 22 · Accepted Answer · answered May 14 '14 at 19:33

22

The clue is that you are calling Math.random which uses a single global instance of Random. So, all your 4 threads compete for the one resource.

Using a thread local Random object will make your execution really parallel:

Random random = new Random();
double[] array = new double[10000000];
for (int i = 0; i < array.length; i++) {
    array[i] = Math.tanh(random.nextDouble());
}

answered May 14 '14 at 19:33

apangin

92,924
10
193
247

Thanks a lot! This was the point, it seems! I fall down to 1700ms. – SamuelHumeau May 14 '14 at 19:35
…and using a thread-local `Random` allows a decent JVM to optimize away the entire operation as it has no global side effect anymore, which will give a *real* speedup once it happened. But after that, parallelization won’t accelerate it further. – Holger May 15 '14 at 08:01

inefficient threads in java

1 Answers1

Linked