The fork/join framework is designed to make divide-and-conquer algorithms easy to parallelize.  More specifically, recursive algorithms where the control path branches out over a few paths and they each process an equal part of the data set.  The typical setup is a new class is created that extends either the RecursiveAction or RecursiveTask class.  The parameters that were sent into the recursive function become member variables in the newly defined class.  Then the recursive calls are replaced by invokeAll(…) rather than the calls to the function itself.

In writing this post, I kept going back for forth on whether I should use Fibonacci numbers as an example or something with more meat to it.  The computations done by each recursive call of a Fibonacci numbers algorithm is too small to matter, not only that, but there are much better non-parallel algorithms for Fibonacci numbers.  In the end, I decided on showing a merge sort.  It is used as the example in the fork/join documentation, but this will be a more complete example showing both the sequential algorithm and the changes made for the parallel version of the algorithm.  You’ll see that it’s not that hard.

First let me start by showing the source code for a typical MergeSort:

public class MergeSort {
 
    private static final int SIZE_THRESHOLD = 16;
 
    public static void sort(Comparable[] a) {
        sort(a, 0, a.length-1);
    }
 
    public static void sort(Comparable[] a, int lo, int hi) {
        if (hi - lo < SIZE_THRESHOLD) {
            insertionsort(a, lo, hi);
            return;
        }
 
        Comparable[] tmp = new Comparable[((hi - lo) / 2) + 1];
        mergeSort(a, tmp, lo, hi);
    }
 
    private static void mergeSort(Comparable[] a, Comparable[] tmp, int lo, int hi) {
        if (hi - lo < SIZE_THRESHOLD) {
            insertionsort(a, lo, hi);
            return;
        }
 
        int m = (lo + hi) / 2;
        mergeSort(a, tmp, lo, m);
        mergeSort(a, tmp, m + 1, hi);
        merge(a, tmp, lo, m, hi);
    }
 
    private static void merge(Comparable[] a, Comparable[] b, int lo, int m, int hi) {
        if (a[m].compareTo(a[m+1]) <= 0)
            return;
 
        System.arraycopy(a, lo, b, 0, m-lo+1);
 
        int i = 0;
        int j = m+1;
        int k = lo;
 
        // copy back next-greatest element at each time
        while (k < j && j <= hi) {
            if (b[i].compareTo(a[j]) <= 0) {
                a[k++] = b[i++];
            } else {
                a[k++] = a[j++];
            }
        }
 
        // copy back remaining elements of first half (if any)
        System.arraycopy(b, i, a, k, j-k);
 
    }
 
    private static void insertionsort(Comparable[] a, int lo, int hi) {
        for (int i = lo+1; i <= hi; i++) {
            int j = i;
            Comparable t = a[j];
            while (j > lo && t.compareTo(a[j - 1]) < 0) {
                a[j] = a[j - 1];
                --j;
            }
            a[j] = t;
        }
    }
}

Now here is the code for the parallel version of MergeSort:

public class ParallelMergeSort {
 
    private static final ForkJoinPool threadPool = new ForkJoinPool();
    private static final int SIZE_THRESHOLD = 16;
 
    public static void sort(Comparable[] a) {
        sort(a, 0, a.length-1);
    }
 
    public static void sort(Comparable[] a, int lo, int hi) {
        if (hi - lo < SIZE_THRESHOLD) {
            insertionsort(a, lo, hi);
            return;
        }
 
        Comparable[] tmp = new Comparable[a.length];
        threadPool.invoke(new SortTask(a, tmp, lo, hi));
    }
 
    /**
     * This class replaces the recursive function that was
     * previously here.
     */
    static class SortTask extends RecursiveAction {
        Comparable[] a;
        Comparable[] tmp;
        int lo, hi;
        public SortTask(Comparable[] a, Comparable[] tmp, int lo, int hi) {
            this.a = a;
            this.lo = lo;
            this.hi = hi;
            this.tmp = tmp;
        }
 
        @Override
        protected void compute() {
            if (hi - lo < SIZE_THRESHOLD) {
                insertionsort(a, lo, hi);
                return;
            }
 
            int m = (lo + hi) / 2;
            // the two recursive calls are replaced by a call to invokeAll
            invokeAll(new SortTask(a, tmp, lo, m), new SortTask(a, tmp, m+1, hi));
            merge(a, tmp, lo, m, hi);
        }
    }
 
    private static void merge(Comparable[] a, Comparable[] b, int lo, int m, int hi) {
        if (a[m].compareTo(a[m+1]) <= 0)
            return;
 
        System.arraycopy(a, lo, b, lo, m-lo+1);
 
        int i = lo;
        int j = m+1;
        int k = lo;
 
        // copy back next-greatest element at each time
        while (k < j && j <= hi) {
            if (b[i].compareTo(a[j]) <= 0) {
                a[k++] = b[i++];
            } else {
                a[k++] = a[j++];
            }
        }
 
        // copy back remaining elements of first half (if any)
        System.arraycopy(b, i, a, k, j-k);
 
    }
 
    private static void insertionsort(Comparable[] a, int lo, int hi) {
        for (int i = lo+1; i <= hi; i++) {
            int j = i;
            Comparable t = a[j];
            while (j > lo && t.compareTo(a[j - 1]) < 0) {
                a[j] = a[j - 1];
                --j;
            }
            a[j] = t;
        }
    }
}

As you can see the majority of the algorithm has remained intact.  As stated above a new class is created that extends RecursiveAction, and the parameters of the function are then passed into that class during creation.  One thing to take note, is that previously only half the size of the original array was created as secondary storage.  Now the entire length of the array is created as a temporary storage.  This is used to avoid different threads needing the same area of the array at the same time.

Changes to the algorithm may be needed, but it definitely helps in making it easier to move to parallel processing.  One other thing to note is the presence of the ForkJoinPool.  The default constructor looks at the processor and determines the appropriate level of parallelism for the task.

I have a quad core CPU, so the ForkJoinPool will spawn at least four threads if necessary. That said, I’ve seen in where only two threads are spawned because more than that was not necessary for the given task. The ForkJoinPool spawns more threads as deemed necessary without starting right at the maximum.

A complete API for the fork/join framework can be found here at the Concurrency JSR-166 Interest Site.  All that is needed for Java 6 is the jsr166y package.

Some other algorithms that are suited for parallelism that I’ve been thinking about are graph searching algorithms such as depth first and breadth first search.  Depending on whether they are done on a tree or a graph determines how much the underlying data structure will need to be changed to support the parallelism.  I plan to look at making a parallel version of the quicksort algorithm using this framework.  Most divide and conquer algorithms can be adapted fairly easily to be multi-threaded using this method, but remember for a performance benefit to be seen the task must be sufficiently large.

The partition function needs to perform the minimal number of swaps possible.  So moving towards the center from both ends and only swapping when both are out of order is the best approach I’ve found so far.  When grouping duplicate elements, they are swapped to the beginning of the partition area as they are found.  Then at the end, a pass is run to move them to their correct location in the final list.  Then instead of returning one number from the partition function, it returns two.  It returns the minimum and maximum indices on the range that has the pivot value.

Another area that I was able to get some performance gain out of was getting rid of the shell sort form the first algorithm.  While that was there to make sure the quicksort did not recurse too deeply, in practice the shell sort algorithm doesn’t run.

Results

Here are the results of JDK 6 MergeSort, Tim Sort, QSort, QSortv2, and Dual Pivot sort 2 benchmarked on the same set of files.  Overall, the new version doesn’t outperform the old version, but I thought it was worth posting my findings.  On most data sets with duplicates it does perform better.  I ran these benchmarks on OpenJDK 7 because I was curious as to how they would compare to one another.

It’s important to note that the tables are speedup relative the Java implementation on the given JDK.  The graphs are the average runtimes for each algorithm.  The reason for doing the average runtime is that it could show the performance difference between Sun’s JDK 6 and OpenJDK 7 build 73.

Sun JDK 6 without Warmup

OpenJDK 7 without Warmup

Sun JDK 6 1000 Warmup Iterations

OpenJDK 7 1000 Warmup Iterations

Sun JDK 6 vs OpenJDK 7 without Warmup

Sun JDK 6 vs OpenJDK 7 1000 Iterations of Warmup

Conclusions

Overall the new version of the Qsort implementation doesn’t improve greatly over the previous implementation.  While it didn’t work out to be the performace improvement I was looking for.  I think the last graph with 1000 iterations of warmup for each algorithm is the most interesting.  The Qsort v2 implementation apparently doesn’t get handled any better by OpenJDK 7.  The partition function is larger after my changes, so perhaps it didn’t JIT very well.  What is interesting is the boost that Tim Sort saw with the change of JDK’s.  Running these benchmarks made me realize that upgrading my Java Runtime will increase the performance of all my Java applications.  It will be interesting to see if the performance carries over to Netbeans and Eclipse; I expect it will.

JDK 6 Sort

The JDK6 implements a fairly standard Merge Sort.  It will switch to an insertion sort at a specific depth.

QSort

This is the implementation of quicksort I outlined in the earlier blog post.  It performed admirably at the time, but how will it hold up against tougher competition.  It’s pretty much an iterative quicksort, that short-circuits to a shell sort if it’s going too deep.

Original QSort Post:
Sort Optimization

Tim Sort

This is an optimized in place variation of a merge sort.  Tim Peters developed this sorting algorithm for the Python programming language.  It is in use by Python and will be used by Java starting with JDK 7.  It takes advantage of partially sorted parts of the list.

Available here:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6804124
http://hg.openjdk.java.net/jdk7/tl/jdk/rev/bfd7abda8f79

Dual Pivot Quick sort

This is a newcomer to the sorting table.  Developed by Vladimir Yaroslavskiy for the inclusion into the Java language.  The premise is the same as quick sort, only it will choose two pivot points rather than one.  He did a full writeup detailing the algorithm, and its benefits.  I did modify it to take the comparable interface, and Vladimir explicitly said this was not the intended target of the algorithm.  He has stated it is designed to work directly with primitive types.  I don’t see how doing an int comparison vs a Integer.compareTo() would be different, as long as they are used uniformly between all algorithms.  Since my sorting algorithm works with comparable, as does Tim Sort, I chose to convert this algorithm to use the Comparable interface also.

Available here:
http://article.gmane.org/gmane.comp.java.openjdk.core-libs.devel/2628

Results

These tables show the speedup relative to JDK 6 with and without warm up.

sorting algorithm speedup without warm up

sorting algorithm speedup with warm up

Here is the original text data if you’re interested in that.  These are in simple table format.  The columns store the runtime in seconds for each algorithm.  The number in parenthesis is the speedup relative to JDK 6.

Without Warmup

With Warmup

Tim Sort is definitely the way to go if you’re interested in a stable sorting algorithm.  I was pretty amazed when I first looked at the results with how well it actually it did.  It really takes advantage of any presorted parts of the lists.  Overall, I’d say my optimized quicksort does fairly well, but maybe it could do better.  I may have to look into that again.

Being most familiar with Java and Ruby, Python seems a little in between, but one feature that stuck out was the regular expression syntax.  Python let’s you put a ‘r’ in front of the quote to denote raw input.  This means you don’t have to escape back slashes twice (ala Java).

For those not familiar with regular expressions, here is an example of a regular expression in a few languages:

Java:
Find Slashes: "[\\\\/]"

Python:
Find Slashes: r"[\\/]"

Ruby:
Find Slashes: /[\\\/]/

This is just a simple example to illustrate a point, but look at Java.  Find slashes has 4 backslashes.  Even if regular expressions were typically this simple (they’re not), that seems unnecessary.  Anytime it’s necessary for a backslash to make it to the regular expression processor, there much be two in the regular expression.  Another quick example, to find a period (.) the regular expression would be “\\.”.  This compounds very quickly, and makes it painful to use regular expressions in Java.

I can’t understand why Java wouldn’t adopt the python syntax of putting a ‘r’ in from of a string to denote raw input.  At quick glance, it doesn’t seem as though it would break any currently in use regular expressions because the old syntax would continue working as expected.  It’d make them better for the future.

Here are some Regular Expression resources that I find useful:

http://www.rubular.com/ A Ruby Regular Expression Tester

http://www.fileformat.info/tool/regex.htm A Java Regular Expression Tester