Multihreading seems to run slower than multiple JVM

Question

Hello,
I have a 4-core machine and I use windows XP.
I build a java application that can launch a thread doing a series of heaving calculation.
The thread takes 8sec to finish and the task manager indicate that 30% of CPU is in use.
If I start 2 threads at the same time, each one finish in 14sec, using 40% of CPU.
Now if I start 2 times my application, and run simultaneously 1 thread in each one, they all finish in 8 sec but using 60% of CPU.
My question is why multithreading goes slower than running 2 JVM in parallel ? When running a profiler on the app, it doesn't seems to wait for shared ressources.
Regarding the CPU usage, it seems that using one JVM doesn't use full available CPU.
Is that a 'normal' behaviour, have you some explanation to it ?
I hope I'm clear enough !
Raphael

Answer 1

I have digged it, and found the part of code that has this behaviour : intialization or copy of a big array.
There is a piece of code showing this behaviour :
public class TestParallel {
     public TestParallel() throws Exception {     
          Thread T1 = new Thread(new Worker());
          Thread T2 = new Thread(new Worker());
          Thread T3 = new Thread(new Worker());
          Thread T4 = new Thread(new Worker());
          long time = System.currentTimeMillis();
          T1.start();
          T1.join();
          T2.start();
          T2.join();
          System.out.println("Serial: " + (System.currentTimeMillis()-time) + " millis");
          time = System.currentTimeMillis();
          T3.start();
          T4.start();
          T3.join();
          T4.join();
          System.out.println("Parallel : " + (System.currentTimeMillis()-time) + " millis");
     public class Worker implements Runnable {
          public void run(){
               long time = System.currentTimeMillis();
               for (int i=0 ; i < 1000 ; i++){
                    double[] aBigArray = new double[500000];
               System.out.println("Worker time : " + (System.currentTimeMillis()-time) + " millis");
     public static void main(String args[]){
          try{
               new TestParallel();
          }catch (Exception e){
               e.printStackTrace();
}I imagine it's a normal behaviour, and there is no workaround.
So I try another way, without array initialization but using System.arrayCopy. But I have the same problem.
You can try by changing Worker runnable this way :
     public class Worker implements Runnable {
          public void run(){
               double[] array1 = new double[500000];
               double[] array2 = new double[500000];
               Arrays.fill(array1, 0.03);
               long time = System.currentTimeMillis();
               for (int i=0 ; i < 1000 ; i++){                    
                    System.arraycopy(array1, 0, array2, 0, 500000);
               System.out.println("Worker time : " + (System.currentTimeMillis()-time) + " millis");
     }I have this output :
Worker time : 3625 millis
Worker time : 3609 millis
Serial: 7266 millis
Worker time : 6531 millis
Worker time : 6563 millis
Parallel : 6594 millisSo two parallel thread working on different arrays with System.arraycopy take nearly twice as much time ???
Edited by: RaphaelH on Jun 23, 2009 4:27 AM