Question

I was trying out the Java ForkJoin framework and wrote a simple test program that sets the pixels of an image to random colors. E.g. it generates pseudo-noise.

But while testing performance I found that it’s actually faster to run single threaded than to run it with multiple threads. I make it run single threaded by passing a high threshold.

This is the class worker class:

public class Noise extends RecursiveAction {

    private BufferedImage image;
    private int xMin;
    private int yMin;
    private int xMax;
    private int yMax;
    private int threshold = 2000000; // max pixels per thread

    public Noise(BufferedImage image, int xMin, int yMin, int xMax, int yMax, int threshold) {
        this.image = image;
        this.xMin = xMin;
        this.yMin = yMin;
        this.xMax = xMax;
        this.yMax = yMax;
        this.threshold = threshold;
    }

    public Noise(BufferedImage image, int xMin, int yMin, int xMax, int yMax) {
        this.image = image;
        this.xMin = xMin;
        this.yMin = yMin;
        this.xMax = xMax;
        this.yMax = yMax;
    }

    @Override
    protected void compute() {
        int ppt = (xMax - xMin) * (yMax - yMin); // pixels pet thread
        if(ppt > threshold) {
            // split
            int verdeling = ((xMax - xMin) / 2) + xMin;
            invokeAll(new Noise(image, xMin, yMin, verdeling, yMax),
                    new Noise(image, verdeling+1, yMin, xMax, yMax));
        }
        else {
            // execute!
            computeDirectly(xMin, yMin, xMax, yMax);
        }
    }

    private void computeDirectly(int xMin, int yMin, int xMax, int yMax) {
        Random generator = new Random();
        for (int x = xMin; x < xMax; x++) {
            for (int y = yMin; y < yMax; y++) {
                //image.setPaint(new Color(generator.nextInt()));
                int rgb = generator.nextInt();
                int red = (rgb >> 16) & 0xFF;
                int green = (rgb >> 8) & 0xFF;
                int blue = rgb & 0xFF;

                red = (int) Math.round((Math.log(255L) / Math.log((double) red)) * 255);
                green = (int) Math.round((Math.log(255L) / Math.log((double) green)) * 255);
                blue = (int) Math.round((Math.log(255L) / Math.log((double) blue)) * 255);

                int rgbSat = red;
                rgbSat = (rgbSat << 8) + green;
                rgbSat = (rgbSat << 8) + blue;

                image.setRGB(x, y, rgbSat);
            }

        }
        Graphics2D g2D = image.createGraphics();
        g2D.setPaint(Color.RED);
        g2D.drawRect(xMin, yMin, xMax-xMin, yMax-yMin);
    }
}

When generating a 6000 * 6000 image the results are:
Single thread: 9.4sec @ 25% CPU load
Multi thread: 16.5sec @ 80%-90% CPU load
(Core2quad Q9450)

Why is the multi-threaded version slower?
How do I fix this?

Answer 1

First of all, F/J is a niche product. If you don’t have a HUGE array and process it as a DAG then you’re using the wrong product. Sure, F/J can make use of multiple processors, but so can just using a simple multi-threaded approach without all the overhead of F/J.

Try using four threads and just give each a quarter of the work directly.

This is the way F/J was meant to be used:

Sum left  = new Sum(array, low, mid);
Sum right = new Sum(array, mid, high);
left.fork();
long rightAns = right.compute();
long leftAns   = left.join();
return leftAns + rightAns;

When you don’t walk down the leaves of a structured tree, then all bets are off.