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


Method Categories _size Mean Error StdDev Median Scaled ScaledSD
Any List 1000 38.88 ns 0.8736 ns 2.1430 ns 38.32 ns 1.00 0.00
Custom List 1000 16.49 ns 0.1061 ns 0.0941 ns 16.50 ns 0.43 0.02
Any Array 1000 27.46 ns 0.2219 ns 0.1732 ns 27.39 ns 1.00 0.00
Custom Array 1000 51.21 ns 1.0657 ns 1.3858 ns 50.60 ns 1.87 0.05
Any HashSet 1000 39.87 ns 0.4526 ns 0.4234 ns 39.89 ns 1.00 0.00
Custom HashSet 1000 15.06 ns 0.1276 ns 0.1193 ns 15.01 ns 0.38 0.00
Any Dictinary 1000 51.14 ns 1.0330 ns 1.4139 ns 50.83 ns 1.00 0.00
Custom Dictinary 1000 17.41 ns 0.2392 ns 0.2238 ns 17.36 ns 0.34 0.01
Any Stack 1000 42.10 ns 0.8651 ns 1.3722 ns 42.29 ns 1.00 0.00
Custom Stack 1000 49.41 ns 1.0303 ns 1.9602 ns 48.60 ns 1.17 0.06
Any Queue 1000 44.68 ns 0.9345 ns 1.6367 ns 44.41 ns 1.00 0.00
Custom Queue 1000 50.68 ns 0.1254 ns 0.0979 ns 50.66 ns 1.14 0.04
Any SortedList 1000 42.89 ns 0.9397 ns 1.8982 ns 42.55 ns 1.00 0.00
Custom SortedList 1000 17.48 ns 0.3887 ns 0.6600 ns 17.17 ns 0.41 0.02
Any SortedSet 1000 220.19 ns 4.3996 ns 4.3210 ns 219.66 ns 1.00 0.00
Custom SortedSet 1000 18.46 ns 0.3891 ns 0.3250 ns 18.34 ns 0.08 0.00

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

Method Categories _size Mean Error StdDev Scaled ScaledSD
Any List 1000 37.52 ns 0.6950 ns 0.6501 ns 1.00 0.00
Custom List 1000 19.69 ns 0.2409 ns 0.2254 ns 0.53 0.01
Any Array 1000 27.11 ns 0.5799 ns 1.0004 ns 1.00 0.00
Custom Array 1000 18.87 ns 0.0513 ns 0.0455 ns 0.70 0.02
Any HashSet 1000 39.42 ns 0.3633 ns 0.3034 ns 1.00 0.00
Custom HashSet 1000 18.76 ns 0.0440 ns 0.0367 ns 0.48 0.00
Any Dictinary 1000 50.13 ns 0.8848 ns 0.8276 ns 1.00 0.00
Custom Dictinary 1000 19.95 ns 0.1613 ns 0.1509 ns 0.40 0.01
Any Stack 1000 39.56 ns 0.5305 ns 0.4702 ns 1.00 0.00
Custom Stack 1000 31.66 ns 0.1149 ns 0.1075 ns 0.80 0.01
Any Queue 1000 42.90 ns 0.2227 ns 0.2083 ns 1.00 0.00
Custom Queue 1000 31.69 ns 0.1502 ns 0.1331 ns 0.74 0.00
Any SortedList 1000 40.53 ns 0.1220 ns 0.1019 ns 1.00 0.00
Custom SortedList 1000 19.71 ns 0.0362 ns 0.0339 ns 0.49 0.00
Any SortedSet 1000 219.88 ns 0.5931 ns 0.4289 ns 1.00 0.00
Custom SortedSet 1000 20.68 ns 0.0420 ns 0.0328 ns 0.09 0.00

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:

Method _size Mean Error StdDev Scaled
Any 1000 230.64 ns 4.5659 ns 6.9727 ns 1.00
Custom 1000 22.23 ns 0.4906 ns 0.7036 ns 0.10
Any 10000 288.27 ns 1.3822 ns 1.0791 ns 1.00
Custom 10000 21.72 ns 0.1599 ns 0.1417 ns 0.08
Any 100000 343.54 ns 6.7375 ns 9.8757 ns 1.00
Custom 100000 21.70 ns 0.3254 ns 0.3043 ns 0.06

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

Method Categories _size Mean Error StdDev Scaled ScaledSD
Any ConcurrentBag 1000 7,065.75 ns 37.9584 ns 33.6491 ns 1.00 0.00
Custom ConcurrentBag 1000 267.89 ns 5.3096 ns 4.7068 ns 0.04 0.00
Any ConcurrentDictionary 1000 39.29 ns 0.7618 ns 0.5948 ns 1.00 0.00
Custom ConcurrentDictionary 1000 6,319.42 ns 27.8553 ns 26.0558 ns 160.88 2.45
Any ConcurrentStack 1000 28.08 ns 0.1495 ns 0.1399 ns 1.00 0.00
Custom ConcurrentStack 1000 3,179.18 ns 23.4161 ns 21.9035 ns 113.23 0.93
Any ConcurrentQueue 1000 72.12 ns 1.4450 ns 1.9779 ns 1.00 0.00
Custom ConcurrentQueue 1000 48.28 ns 0.3934 ns 0.3487 ns 0.67 0.02

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[index];
  return num;
}

And it takes more time in collection with more elements:

Method _size Mean Error StdDev Scaled ScaledSD
Any 10 37.79 ns 0.7850 ns 1.0207 ns 1.00 0.00
Custom 10 239.63 ns 2.2313 ns 2.0872 ns 6.34 0.18
Any 100 34.46 ns 0.4621 ns 0.4322 ns 1.00 0.00
Custom 100 823.62 ns 16.0617 ns 17.8525 ns 23.90 0.58
Any 1000 38.22 ns 0.5311 ns 0.4968 ns 1.00 0.00
Custom 1000 6,264.17 ns 15.1371 ns 12.6402 ns 163.92 2.09
Any 10000 35.57 ns 0.5368 ns 0.5021 ns 1.00 0.00
Custom 10000 24,819.31 ns 283.1502 ns 264.8588 ns 697.82 11.97

Profiler shows that locking takes a lot of time:

No implementation

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.

Links

blog comments powered by Disqus