A look at STL/CLR performance for linear containers

Introduction

This article is an attempt to compare the general performance of STL/CLR's sequence containers with the .NET generic List<T> collection class. Before I began work on the article, I strongly believed that the STL/CLR containers would be yards faster. To my utmost surprise, I found that this was not so and that List<T> surpassed the STL/CLR collections with ease.

How I compared performance

I wanted to keep things simple and used the common technique of repeating a specific operation several times. To smoothen the design, I have an interface as follows :-

namespace STLCLRTests 
{
    public interface class IMeasurable 
    {
        Int64 RunCode(int iterations);
    };
}

RunCode would run a specific piece of code as many times as specified by iterations, and would return the time taken in milliseconds. And I have the following abstract class that implements this interface.

namespace STLCLRTests 
{
    public ref class MeasurableDoubleOp abstract : IMeasurable
    {
    private:
        static Stopwatch^ stopWatch = gcnew Stopwatch();

    public:
        virtual Int64 RunCode(int iterations)
        {
            Initialize();

            stopWatch->Reset();
            stopWatch->Start();

            RunCodeFirstOp(iterations);
            RunCodeSecondOp(iterations);

            stopWatch->Stop();

            return stopWatch->ElapsedMilliseconds;
        }

    protected:
        virtual void Initialize() {}
        virtual void RunCodeFirstOp(int iterations) abstract;
        virtual void RunCodeSecondOp(int iterations) abstract;
    };
}

To profile a certain collection class, I just derive from this abstract class and implement RunCodeFirstOp and RunCodeSecondOp. I also have a MeasurableSingleOp class for doing tests that do not involve a two-step operation.

STL vector vs List<T> - basic insertion/removal

Here are the implementations of the vector specific and List<T> specific classes.

namespace STLCLRTests 
{
    public ref class VectorInsertRemove : MeasurableDoubleOp
    {
    private:
        cliext::vector<int> vector;

    protected:
        IEnumerable<int>^ GetVector()
        {
            return %vector;
        }

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                vector.push_back(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            for(int count=0; count<iterations; count++)
            {
                vector.pop_back();
            }
        }
    };
}
namespace STLCLRTests 
{
    public ref class GenericListInsertRemove : MeasurableDoubleOp
    {
    private:
        List<int> list;

    protected:
        IEnumerable<int>^ GetList()
        {
            return %list;
        }

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                list.Add(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            for(int count=0; count<iterations; count++)
            {
                list.RemoveAt(list.Count - 1);
            }
        }
    };
}

And here are my test results. As you can see, the BCL class (List<T>) completely outperformed the STL/CLR vector class.

Here's a graphical plot of how the two containers performed. Clearly, the BCL class's performance was quite superior to the STL vector's.

As you can imagine I was quite surprised by this result. Just for the heck of it I thought I should also compare the standard STL vector with the STL/CLR vector implementation. Note than I am still using managed code (/clr) - the standard STL code is also compiled as /clr. Here are my surprising results.

Based on that result, you should absolutely avoid compiling native STL code using /clr. Merely porting to STL/CLR would help performance in a big way. You might find that all you need is a namespace change (cliext to std) and you may not have to change too much code elsewhere. And no, I did not conclude this merely on my test results with vector, I compared the standard list and the STL/CLR list containers with the following results.

As you can see, the difference in performance is non-trivial. Please note that we are not comparing the native performance of STL here. We are comparing how the native implementation when compiled under /clr compares with the CLR implementation of STL.

STL list vs List<T> - basic insertion/removal

Here's my implementation for the STL list specific class.

namespace STLCLRTests 
{
    public ref class StlListInsertRemove : MeasurableDoubleOp
    {
    private:
        cliext::list<int> list;

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                list.push_back(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            for(int count=0; count<iterations; count++)
            {
                list.pop_back();
            }
        }
    };
}

And here are my test results. Here, the contrast is even more - not surprising really, as the STL list will always be slower than the STL vector for straight inserts and removals.

And here's a graphical plot of the results.

STL deque vs List<T> - basic insertion/removal

Here's the deque implementation.

namespace STLCLRTests 
{
    public ref class DequeInsertRemove : MeasurableDoubleOp
    {
    private:
        cliext::deque<int> deque;

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                deque.push_back(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            for(int count=0; count<iterations; count++)
            {
                deque.pop_back();
            }
        }
    };
}

Here are my results. Nothing's changed in the pattern - the BCL class is way faster here too.

And here's the graph.

The BCL equivalent of a queue is the Queue<T> class - so just to be sure we are comparing apples with apples, I went ahead and ran tests comparing the STL/CLR deque with the BCL Queue<T>. My results and the corresponding graph follow.

The Queue<T> class seems to be marginally slower than List<T> but is still way faster than the STL/CLR deque container.

STL vector vs List<T> - basic iteration

This time, I wanted to test the speed with which we can iterate over a linear collection. Here are the vector and List<T> specific iteration test implementations.

namespace STLCLRTests 
{
    public ref class VectorIterate : MeasurableDoubleOp
    {
    private:
        cliext::vector<int> vector;

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            vector.clear();

            for(int count=0; count<iterations; count++)
            {
                vector.push_back(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            for(cliext::vector<int>::iterator it = vector.begin(); it != vector.end(); it++)
            {
            }
        }
    };
}
namespace STLCLRTests 
{
    public ref class GenericListIterate : MeasurableDoubleOp
    {
    private:
        List<int> list;

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            list.Clear();

            for(int count=0; count<iterations; count++)
            {
                list.Add(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            for each(int x in list)
            {
            }
        }
    };
}

Here are my test results. The results further proved the superior efficiency of the List<T> class.

And here's the corresponding graph.

STL vector vs List<T> - Linq access (where)

For the Linq tests, I used a C# project (for easier syntax). I derived from the insert tester and merely overrode the RunCodeSecondOp method as I wanted to keep the insertion code intact.

namespace LinqTests
{
    public class VectorLinqWhere : VectorInsertRemove
    {
        public override void RunCodeSecondOp(int iterations)
        {
            IEnumerable<int> _vector = GetVector();
            var newVector = _vector.Where(x => x % 2 == 0);
        }
    }
}
namespace LinqTests
{
    public class GenericListLinqWhere : GenericListInsertRemove
    {
        public override void RunCodeSecondOp(int iterations)
        {
            IEnumerable<int> _list = GetList();
            var newList = _list.Where(x => x % 2 == 0);
        }
    }
}

Here are the results of my test runs. The results here are partially contaminated by the fact that the insertion code speed differences would also come into play. But the difference in performance is large enough to safely ignore that for now, and again LINQ works much faster on the BCL class as compared to the STL/CLR version.

And here's the graph.

STL vector vs List<T> - Linq access (take)

This is similar to the previous one except I use Take instead of Where.

namespace LinqTests
{
    public class VectorLinqTake : VectorInsertRemove
    {
        public override void RunCodeSecondOp(int iterations)
        {
            IEnumerable<int> _vector = GetVector();
            var newVector = _vector.Take(_vector.Count() / 2);
        }
    }
}
namespace LinqTests
{
    public class GenericListLinqTake : GenericListInsertRemove
    {
        public override void RunCodeSecondOp(int iterations)
        {
            IEnumerable<int> _list = GetList();
            var newList = _list.Take(_list.Count() / 2);
        }
    }
}

Here's the result of my tests. These results are very similar to the previous test.

And the corresponding graph.

Sorting

I ran tests comparing sorting speeds of the List<T> class with the STL/CLR vector and list containers. The code used follows.

namespace STLCLRTests 
{
    public ref class GenericListSort : MeasurableDoubleOp
    {
    private:
        List<int> list;

    protected:
        IEnumerable<int>^ GetList()
        {
            return %list;
        }

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                list.Add(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            list.Sort();
        }
    };
}

namespace STLCLRTests 
{
    public ref class StlListSort : MeasurableDoubleOp
    {
    private:
        cliext::list<int> list;

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                list.push_back(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            list.sort();
        }
    };
}
namespace STLCLRTests 
{
    public ref class VectorSort : MeasurableDoubleOp
    {
    private:
        cliext::vector<int> vector;

    protected:
        IEnumerable<int>^ GetVector()
        {
            return %vector;
        }

    public:
        virtual void RunCodeFirstOp(int iterations) override
        {
            for(int count=0; count<iterations; count++)
            {
                vector.push_back(10);
            }
        }

        virtual void RunCodeSecondOp(int iterations) override 
        {
            sort(vector.begin(), vector.end());
        }
    };
}

Here are the results for vector versus List<T>.

And here are my results for stl list versus List<T>.

Conclusion

One of the features that was strongly marketed before STL/CLR was released was its performance benefits over regular .NET collections. But the .NET generic List<T> seems to be much faster. At this stage all I can think of as a valid case for using STL/CLR would be when doing a first-level port of existing C++ code ( that heavily uses STL) to managed code.