Edit

Share via

How to: Implement a Partitioner for Static Partitioning

  • Article

The following example shows one way to implement a simple custom partitioner for PLINQ that performs static partitioning. Because the partitioner does not support dynamic partitions, it is not consumable from Parallel.ForEach. This particular partitioner might provide speedup over the default range partitioner for data sources for which each element requires an increasing amount of processing time.

Example

// A static range partitioner for sources that require
// a linear increase in processing time for each succeeding element.
// The range sizes are calculated based on the rate of increase
// with the first partition getting the most elements and the
// last partition getting the least.
class MyPartitioner : Partitioner<int>
{
    int[] source;
    double rateOfIncrease = 0;

    public MyPartitioner(int[] source, double rate)
    {
        this.source = source;
        rateOfIncrease = rate;
    }

    public override IEnumerable<int> GetDynamicPartitions()
    {
        throw new NotImplementedException();
    }

    // Not consumable from Parallel.ForEach.
    public override bool SupportsDynamicPartitions
    {
        get
        {
            return false;
        }
    }

    public override IList<IEnumerator<int>> GetPartitions(int partitionCount)
    {
        List<IEnumerator<int>> _list = new List<IEnumerator<int>>();
        int end = 0;
        int start = 0;
        int[] nums = CalculatePartitions(partitionCount, source.Length);

        for (int i = 0; i < nums.Length; i++)
        {
            start = nums[i];
            if (i < nums.Length - 1)
                end = nums[i + 1];
            else
                end = source.Length;

            _list.Add(GetItemsForPartition(start, end));

            // For demonstration.
            Console.WriteLine("start = {0} b (end) = {1}", start, end);
        }
        return (IList<IEnumerator<int>>)_list;
    }
    /*
     *
     *
     *                                                               B
      // Model increasing workloads as a right triangle           /  |
         divided into equal areas along vertical lines.         / |  |
         Each partition  is taller and skinnier               /   |  |
         than the last.                                     / |   |  |
                                                          /   |   |  |
                                                        /     |   |  |
                                                      /  |    |   |  |
                                                    /    |    |   |  |
                                            A     /______|____|___|__| C
     */
    private int[] CalculatePartitions(int partitionCount, int sourceLength)
    {
        // Corresponds to the opposite side of angle A, which corresponds
        // to an index into the source array.
        int[] partitionLimits = new int[partitionCount];
        partitionLimits[0] = 0;

        // Represent total work as rectangle of source length times "most expensive element"
        // Note: RateOfIncrease can be factored out of equation.
        double totalWork = sourceLength * (sourceLength * rateOfIncrease);
        // Divide by two to get the triangle whose slope goes from zero on the left to "most"
        // on the right. Then divide by number of partitions to get area of each partition.
        totalWork /= 2;
        double partitionArea = totalWork / partitionCount;

        // Draw the next partitionLimit on the vertical coordinate that gives
        // an area of partitionArea * currentPartition.
        for (int i = 1; i < partitionLimits.Length; i++)
        {
            double area = partitionArea * i;

           // Solve for base given the area and the slope of the hypotenuse.
            partitionLimits[i] = (int)Math.Floor(Math.Sqrt((2 * area) / rateOfIncrease));
        }
        return partitionLimits;
    }

    IEnumerator<int> GetItemsForPartition(int start, int end)
    {
        // For demonstration purposes. Each thread receives its own enumerator.
        Console.WriteLine("called on thread {0}", Thread.CurrentThread.ManagedThreadId);
        for (int i = start; i < end; i++)
            yield return source[i];
    }
}

class Consumer
{
    public static void Main2()
    {
        var source = Enumerable.Range(0, 10000).ToArray();

        Stopwatch sw = Stopwatch.StartNew();
        MyPartitioner partitioner = new MyPartitioner(source, .5);

        var query = from n in partitioner.AsParallel()
                    select ProcessData(n);

        foreach (var v in query) { }
        Console.WriteLine("Processing time with custom partitioner {0}", sw.ElapsedMilliseconds);

        var source2 = Enumerable.Range(0, 10000).ToArray();

        sw = Stopwatch.StartNew();

        var query2 = from n in source2.AsParallel()
                    select ProcessData(n);

        foreach (var v in query2) { }
        Console.WriteLine("Processing time with default partitioner {0}", sw.ElapsedMilliseconds);
    }

    // Consistent processing time for measurement purposes.
    static int ProcessData(int i)
    {
        Thread.SpinWait(i * 1000);
        return i;
    }
}

The partitions in this example are based on the assumption of a linear increase in processing time for each element. In the real world, it might be difficult to predict processing times in this way. If you are using a static partitioner with a specific data source, you can optimize the partitioning formula for the source, add load-balancing logic, or use a chunk partitioning approach as demonstrated in How to: Implement Dynamic Partitions.

See also