Fork Join Matrix Multiplication in Java

Question

I’m doing some performance research on the fork/join framework in Java 7. To improve the test results I want to use different recursive algorithms during the tests. One of t

Answer 1

You are accumulating the results in C[cRow + i][cCol + j] += s00; and the like. This is not a thread safe operation so you must synchronize the row or ensure that only one task ever updates the cell. Without this you will see random cells being set incorrectly.

I would check you get the right answer with a concurrency of 1.

BTW: float may not be the best choice here. It has a fairly low number of digits of precision and in heavy matrix operations (which I assume you are doing or there wouldn't be much point using multiple threads) the rounding error can use up most or all your precision. I would suggest considering double instead.

e.g. float has about 7 digits of precision and one rule of thumb is that the error is proportional to the number of calculations. So for a 1K x 1K matrix you might have 4 digits of precision left. For 10K x 10K you might only have three at best. double has 16 digits of precision meaning you may have 12 digits of precision after a 10K x 10K mutlipication.

Answer 2

As you already noticed, sequential execution of subtasks that belong to the same quadrant is important for this algorithm. So, you need to implement your own seq() function, for example, as follows, and use it as in original code:

public ForkJoinTask<?> seq(final ForkJoinTask<?> a, final ForkJoinTask<?> b) {
    return adapt(new Runnable() {
        public void run() {
            a.invoke();
            b.invoke();
        }
    });
}