Sunday, 6 January 2013

Map and Reduce in PLinq

This article will present some code that can be used to apply a Map and a Reduce transformation of a provided input array of a given type of objects, using PLINQ. For developers familiar with LINQ, PLINQ is the parallel version of Linq. First, the source code will be presented, then the code will be explained.

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;

namespace TestParallelMap
{
class Program
{

static void Main(string[] args)
{
var nums = Enumerable.Range(0, 1001).ToArray();
Func<int,TaggedResult<int,double>> mapFunction = n => new TaggedResult<int,double>(n, Math.Pow(n, 2));
var result = ParallelMap(nums, mapFunction);

foreach (var r in result)
{
Console.WriteLine("Got result {0} for the num {1}", r.OutputValue, r.InputtedValue);
}

Console.WriteLine("Press the any key to continue ...");
Console.ReadKey();

var resultSecond = ParallelReduce(nums, 0, (first, second) => first + second);

Console.WriteLine("Got result {0}", resultSecond);

Console.WriteLine("Press the any key to continue ...");
Console.ReadKey();

}

private static TValue ParallelReduce<TValue>(TValue[] source, TValue seedValue, Func<TValue, TValue, TValue> reduceFunction)
{
return source.AsParallel().AsOrdered().Aggregate(
seedValue,
(localResult, nextValue) => reduceFunction(localResult, nextValue),
(overallResult, localResult) => reduceFunction(overallResult, localResult),
overallResult => overallResult);

}

private static TOutput[] ParallelMap<TInput, TOutput>(TInput[] source, Func<TInput, TOutput> mapFunction)
{
return source.AsParallel().AsOrdered().Select(x => mapFunction(x)).ToArray();
}

private class TaggedResult<TInput, TOutput>
{
public TaggedResult(TInput inputtedValue, TOutput outputtedValue)
{
InputtedValue = inputtedValue;
OutputValue = outputtedValue;
}

public TInput InputtedValue { get; set; }

public TOutput OutputValue { get; set; }
}

}
}


The source code presented above is a simple console application. It is written in C# (obviously) and uses PLINQ to prosess the input array. The Map method will transform each value in the input array with a provided mapfunction. The source code uses Func to provide a function pointer to the method that will be executed to produce each item. The generic arguments control the type of the resulting array:


private static TOutput[] ParallelMap<TInput, TOutput>(TInput[] source, Func<TInput, TOutput> mapFunction)
{
return source.AsParallel().AsOrdered().Select(x => mapFunction(x)).ToArray();
}


Note that PLINQ is very similar to LINQ and the main difference is the AsParallel() method. Also note that AsOrdered() will mean that the ordering will be maintained in the output array. If the order does not matter, there is a performance gain by omitting the call to AsOrdered().. Further, note that one uses the Select method of PLINQ to perform a projection where one at the same time performes the map method. The Reduce method is a bit more complicated. It uses PLINQ's Aggregate method, which really is a bit tough of understanding. Let's repeat the source code:

private static TValue ParallelReduce<TValue>(TValue[] source, TValue seedValue, Func<TValue, TValue, TValue> reduceFunction)
{
return source.AsParallel().AsOrdered().Aggregate(
seedValue,
(localResult, nextValue) => reduceFunction(localResult, nextValue),
(overallResult, localResult) => reduceFunction(overallResult, localResult),
overallResult => overallResult);

}


Actually, the PLINQ aggregate method is very versatile and powerful method when you want to process an entire array and perform a reduction method, in our case just a summation, but this could be anything, for example statistical methods and so on. The first argument is the seed value, the next parameter is a lambda expression that will perform the stepwise part of the reduction, the next parameter is the lambda expression that merges the stepwise result with the overall result - and the final fourth parameter is just a simple lambda which tells we do not want to do anything with the final overall result than propagate the value to the processed final result which then is returned. There are many more patterns that are a bit more complicated too, like MapReduce (which combines map and reduce), but this will be a relatively gentle introduction to the functionality that PLINQ presents. It is impressive to think that all the thread handling will be automatically be handled by PLINQ. From relatively simple code one can perform a multitude of different transformation from this source code with Map and Reduce, powered by PLINQ and generics.