Any vs Count

This post started from discussion on SO: Why LINQ method Any does not check Count? After it I decided to check the most common collections on performance in methods Any and Count.

First of all, we create new extension method:

public static class EnumerableExtensions
{
	public static bool CustomAny<TSource>(this IEnumerable<TSource> source)
	{
		if (source == null)
		{
			throw new ArgumentNullException(nameof(source));
		}

		if (source is ICollection<TSource> collection)
		{
			return collection.Count != 0;
		}

		using (var enumerator = source.GetEnumerator())
		{
			if (enumerator.MoveNext())
				return true;
		}

		return false;
	}
}

For benchmarking I use BenchmarkDotNet.

Collections to test:

• List
• Array
• HashSet
• Dictionary
• Stack
• Queue
• SortedList
• SortedSet

I create collections with 1000 elements and test methods Any and CustomAny:

[CategoriesColumn]
[GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory)]
public class SimpleCollections
{
	[Params(1000)]
	public int _size;

	private List<int> _list;
	private int[] _array;
	private HashSet<int> _hashSet;
	private Dictionary<int, int> _dictionary;
	private Stack<int> _stack;
	private Queue<int> _queue;
	private SortedList<int, int> _sortedList;
	private SortedSet<int> _sortedSet;

	[GlobalSetup]
	public void SetUp()
	{
		_list = new List<int>();
		_array = new int[_size];
		_hashSet = new HashSet<int>();
		_dictionary = new Dictionary<int, int>();
		_stack = new Stack<int>();
		_queue = new Queue<int>();
		_sortedList = new SortedList<int, int>();
		_sortedSet = new SortedSet<int>();

		for (int i = 0; i < _size; i++)
		{
			_list.Add(i);
			_array[i] = i;
			_hashSet.Add(i);
			_dictionary[i] = i;
			_stack.Push(i);
			_queue.Enqueue(i);
			_sortedList.Add(i, i);
			_sortedSet.Add(i);
		}
	}
	
	[Benchmark(Baseline = true, Description = "Any"), BenchmarkCategory("List")]
	public bool ListAny()
	{
		return _list.Any();
	}

	[Benchmark(Description = "Custom"), BenchmarkCategory("List")]
	public bool ListCount()
	{
		return _list.CustomAny();
	}
	
	///other collections
}

Benchmarks are grouped by category — type of collection. Method Any is default Enumerable.Any, Custom — our custom implementation EnumerableExtensions.CustomAny.

Any is a baseline benchmark in category.

First run result

BenchmarkDotNet=v0.10.14, OS=Windows 10.0.17134
Intel Core i7-2670QM CPU 2.20GHz (Sandy Bridge), 1 CPU, 8 logical and 4 physical cores
Frequency=2143565 Hz, Resolution=466.5126 ns, Timer=TSC
  [Host]     : .NET Framework 4.7.1 (CLR 4.0.30319.42000), 32bit LegacyJIT-v4.7.3110.0
  DefaultJob : .NET Framework 4.7.1 (CLR 4.0.30319.42000), 32bit LegacyJIT-v4.7.3110.0

Here and later I use legends:

  Categories : Collection type
  _size      : Number of elements in collection
  Mean       : Arithmetic mean of all measurements
  Error      : Half of 99.9% confidence interval
  StdDev     : Standard deviation of all measurements
  Median     : Value separating the higher half of all measurements (50th percentile)
  Scaled     : Mean(CurrentBenchmark) / Mean(BaselineBenchmark)
  ScaledSD   : Standard deviation of ratio of distribution of [CurrentBenchmark] and [BaselineBenchmark]
  1 ns       : 1 Nanosecond (0.000000001 sec)

As we see, in common the new method with Count is faster than the old one or the same, except some collections. Lets examine them in more detail.

Array

Our new method gives much worse results. Why? Answer is not obvious. Array can be converted to ICollection<T>, but it takes much time due to internal implementation of array in .NET.

So, lets add array check:

public static class EnumerableExtensions
{
	public static bool CustomAny<TSource>(this IEnumerable<TSource> source)
	{
		if (source == null)
		{
			throw new ArgumentNullException(nameof(source));
		}

		if (source is TSource[] array)
		{
			return array.Length != 0;
		}

		if (source is ICollection<TSource> collection)
		{
			return collection.Count != 0;
		}

		using (var enumerator = source.GetEnumerator())
		{
			if (enumerator.MoveNext())
				return true;
		}

		return false;
	}
}

Stack and Queue

These collections do not implement ICollection<T> interface:

public class Stack<T> : IEnumerable<T>, IEnumerable, ICollection, IReadOnlyCollection<T>

public class Queue<T> : IEnumerable<T>, IEnumerable, ICollection, IReadOnlyCollection<T>

Only ICollection, we can use it as the most common:

public static class EnumerableExtensions
{
	public static bool CustomAny<TSource>(this IEnumerable<TSource> source)
	{
		if (source == null)
		{
			throw new ArgumentNullException(nameof(source));
		}

		if (source is TSource[] array)
		{
			return array.Length != 0;
		}

		if (source is ICollection<TSource> collection)
		{
			return collection.Count != 0;
		}

		if (source is ICollection baseCollection)
		{
			return baseCollection.Count != 0;
		}

		using (var enumerator = source.GetEnumerator())
		{
			if (enumerator.MoveNext())
				return true;
		}

		return false;
	}
}

Second run result

Ok, in all tests our method is better than default one. But there is one collection, where result is 10 times better — SortedSet.

SortedSet — what’s wrong?

Why this collection is so bad in simple method Any? As we know, Any just create enumerator and tries to take first element. Open source code of SortedSet:

IEnumerator<T> IEnumerable<T>.GetEnumerator()
{
  return (IEnumerator<T>) new SortedSet<T>.Enumerator(this);
}

GetEnumerator creates new struct of type Enumerator, constructor gets SortedSet<T> as a parameter:

internal Enumerator(SortedSet<T> set)
{
  this.tree = set;
  this.tree.VersionCheck();
  this.version = this.tree.version;
  this.stack = new Stack<SortedSet<T>.Node>(2 * SortedSet<T>.log2(set.Count + 1));
  this.current = (SortedSet<T>.Node) null;
  this.reverse = false;
  this.siInfo = (SerializationInfo) null;
  this.Intialize();
}

Here is nothing interesting, open Intialize (misprint in source code) method:

private void Intialize()
{
  this.current = (SortedSet<T>.Node) null;
  SortedSet<T>.Node node1 = this.tree.root;
  while (node1 != null)
  {
    SortedSet<T>.Node node2 = this.reverse ? node1.Right : node1.Left;
    SortedSet<T>.Node node3 = this.reverse ? node1.Left : node1.Right;
    if (this.tree.IsWithinRange(node1.Item))
    {
	  this.stack.Push(node1);
	  node1 = node2;
    }
    else
	  node1 = node2 == null || !this.tree.IsWithinRange(node2.Item) ? node3 : node2;
  }
}

It is worth noting that SortedSet is built on the basis of binary tree, and method Intialize bypasses all nodes in tree and puts elements into the stack. So, when you create enumerator of SortedSet, internal implementation creates new collections with all elements. The more elements you have, the more time it takes:

Ok, we checked simple most popular collections, and I think, we can use Count property instead of Enumerable.Any. But is it safe to use Count every time? Is there a collection, where Count takes more time than creating enumerator?

Concurrent collections

All collections we checked before are not thread-safe, and can not be used in multi-threading applications. There is a special class of collections — concurrent collections:

• ConcurrentBag
• ConcurrentDictionary
• ConcurrentQueue
• ConcurrentStack

I will fill collections from different threads to simulate multi-threading environment:

[GlobalSetup]
public void SetUp()
{
	_bag = new ConcurrentBag<int>();
	_dictionary = new ConcurrentDictionary<int, int>();
	_queue = new ConcurrentQueue<int>();
	_stack = new ConcurrentStack<int>();

	var tasksCount = 10;
	var batch = _size / tasksCount;

	var tasks = new Task[tasksCount];

	for (int i = 0; i < tasksCount; i++)
	{
		var task = i;

		tasks[task] = Task.Run(() =>
		{
			var from = task * batch;
			var to = (task + 1) * batch;

			for (int j = from; j < to; j++)
			{
				_bag.Add(j);
				_dictionary[j] = j;
				_queue.Enqueue(j);
				_stack.Push(j);
			}
		});
	}

	Task.WaitAll(tasks);
}

Concurrent collections result

We have got different unpredictable results. Somewhere Any is very slow, and vice versa.

ConcurrentBag

Method GetEnumerator of ConcurrentBag works like SortedSet: it locks collection and copies all elements into new list:

private List<T> ToList()
{
  List<T> objList = new List<T>();
  for (ConcurrentBag<T>.ThreadLocalList threadLocalList = this.m_headList; threadLocalList != null; threadLocalList = threadLocalList.m_nextList)
  {
	for (ConcurrentBag<T>.Node node = threadLocalList.m_head; node != null; node = node.m_next)
	  objList.Add(node.m_value);
  }
  return objList;
}

Count property locks collection and just sum up all internal lists counts:

private int GetCountInternal()
{
  int num = 0;
  for (ConcurrentBag<T>.ThreadLocalList threadLocalList = this.m_headList; threadLocalList != null; threadLocalList = threadLocalList.m_nextList)
	checked { num += threadLocalList.Count; }
  return num;
}

ConcurrentDictionary

GetEnumerator method does not create new collection and uses yield to iterate through collection, so, Any is cheap:

public IEnumerator<KeyValuePair<TKey, TValue>> GetEnumerator()
{
  foreach (ConcurrentDictionary<TKey, TValue>.Node bucket in this.m_tables.m_buckets)
  {
	ConcurrentDictionary<TKey, TValue>.Node current;
	for (current = Volatile.Read<ConcurrentDictionary<TKey, TValue>.Node>(ref bucket); current != null; current = current.m_next)
	  yield return new KeyValuePair<TKey, TValue>(current.m_key, current.m_value);
	current = (ConcurrentDictionary<TKey, TValue>.Node) null;
  }
}

Count locks collection and sums up counts of internal tables:

public int Count
{
  [__DynamicallyInvokable] get
  {
	int locksAcquired = 0;
	try
	{
	  this.AcquireAllLocks(ref locksAcquired);
	  return this.GetCountInternal();
	}
	finally
	{
	  this.ReleaseLocks(0, locksAcquired);
	}
  }
}

private int GetCountInternal()
{
  int num = 0;
  for (int index = 0; index < this.m_tables.m_countPerLock.Length; ++index)
	num += this.m_tables.m_countPerLock;
  return num;
}

And it takes more time in collection with more elements:

Profiler shows that locking takes a lot of time:

ConcurrentStack

GetEnumerator uses lazy enumeration and it’s cheap to use:

private IEnumerator<T> GetEnumerator(ConcurrentStack<T>.Node head)
{
  for (ConcurrentStack<T>.Node current = head; current != null; current = current.m_next)
	yield return current.m_value;
}

But Count just counts all elements even without locking:

public int Count
{
  [__DynamicallyInvokable] get
  {
	int num = 0;
	for (ConcurrentStack<T>.Node node = this.m_head; node != null; node = node.m_next)
	  ++num;
	return num;
  }
}

ConcurrentQueue

GetEnumerator implementation is a bit complicated, but it uses lazy yield internal. Count uses simple calculations without enumerating all elements. You can use both methods.

As you see, in concurrent world everything is not so simple and internal implementation differs very much.

Universal method for everything?

No, it’s impossible. You can use custom Any method that checks Count property for some simple collections, but better use Count or Any depending on collection.