爲了讓共享的數組,集合可以被多線程更新,咱們如今(.net4.0以後)可使用併發集合來實現這個功能。而System.Collections和System.Collections.Generic命名空間中所提供的經典列表,集合和數組都不是線程安全的,若是要使用,還須要添加代碼來同步。編程
先看一個例子,經過並行循環向一個List<string>集合添加元素。由於List不是線程安全的,因此必須對Add方法加鎖來串行化。api
任務開始:數組
private static int NUM_AES_KEYS =80000; static void Main(string[] args) { Console.WriteLine("任務開始..."); var sw = Stopwatch.StartNew(); for (int i = 0; i < 4; i++) { ParallelGennerateMD5Keys(); Console.WriteLine(_keyList.Count); } Console.WriteLine("結束時間:" + sw.Elapsed); Console.ReadKey(); }
private static List<string> _keyList; private static void ParallelGennerateMD5Keys() { _keyList=new List<string>(NUM_AES_KEYS); Parallel.ForEach(Partitioner.Create(1, NUM_AES_KEYS + 1), range => { var md5M = MD5.Create(); for (int i = range.Item1; i < range.Item2; i++) { byte[] data = Encoding.Unicode.GetBytes(Environment.UserName + i); byte[] result = md5M.ComputeHash(data); string hexString = ConverToHexString(result); lock (_keyList) { _keyList.Add(hexString); } } }); }
但若是咱們去掉lock,獲得的結果以下: 安全
沒有一次是滿80000的。lock關鍵字建立了一個臨界代碼區,當一個任務進入以後,其餘任務會被阻塞並等待進入。lock關鍵字引入了必定的開銷,並且會下降可擴展性。對於這個問題,.Net4.0提供了System.Collections.Concurrent命名空間用於解決線程安全問題,它包含了5個集合:ConcurrentQueue<T>,ConcurrentStack<T>,ConcurrentBag<T>,BlockingCollection<T>,ConcurrentDictionary<TKey,TValue>。這些集合都在某種程度上使用了無鎖技術,性能獲得了提高。數據結構
ConcurrentQueue多線程
一個FIFO(先進先出)的集合。支持多任務進併發行入隊和出隊操做。 併發
ConcurrentQueue是徹底無鎖的,它是System.Collections.Queue的併發版本。提供三個主要的方法:dom
修改上面的代碼:ide
private static ConcurrentQueue<string> _keyQueue; private static void ParallelGennerateMD5Keys() { _keyQueue = new ConcurrentQueue<string>(); Parallel.ForEach(Partitioner.Create(1, NUM_AES_KEYS + 1), range => { var md5M = MD5.Create(); for (int i = range.Item1; i < range.Item2; i++) { byte[] data = Encoding.Unicode.GetBytes(Environment.UserName + i); byte[] result = md5M.ComputeHash(data); string hexString = ConverToHexString(result); _keyQueue.Enqueue(hexString); } }); }
結果以下:函數
能夠看見,它的使用很簡單,不用擔憂同步問題。接下咱們經過生產者-消費者模式,對上面的問題進行改造,分解成兩個任務。使用兩個共享的ConcurrentQueue實例。_byteArraysQueue 和 _keyQueue ,ParallelGennerateMD5Keys 方法生產byte[],ConverKeysToHex方法去消費併產生key。
private static ConcurrentQueue<string> _keyQueue; private static ConcurrentQueue<byte[]> _byteArraysQueue; private static void ParallelGennerateMD5Keys(int maxDegree) { var parallelOptions = new ParallelOptions{MaxDegreeOfParallelism = maxDegree}; var sw = Stopwatch.StartNew(); _keyQueue = new ConcurrentQueue<string>(); Parallel.ForEach(Partitioner.Create(1, NUM_AES_KEYS + 1),parallelOptions, range => { var md5M = MD5.Create(); for (int i = range.Item1; i < range.Item2; i++) { byte[] data = Encoding.Unicode.GetBytes(Environment.UserName + i); byte[] result = md5M.ComputeHash(data); _byteArraysQueue.Enqueue(result); } }); Console.WriteLine("MD5結束時間:" + sw.Elapsed); } private static void ConverKeysToHex(Task taskProducer) { var sw = Stopwatch.StartNew(); while (taskProducer.Status == TaskStatus.Running || taskProducer.Status == TaskStatus.WaitingToRun || _byteArraysQueue.Count > 0) { byte[] result; if (_byteArraysQueue.TryDequeue(out result)) { string hexString = ConverToHexString(result); _keyQueue.Enqueue(hexString); } } Console.WriteLine("key結束時間:" + sw.Elapsed); }
此次我修改了執行次數爲180000
private static int NUM_AES_KEYS =180000; static void Main(string[] args) { Console.WriteLine("任務開始..."); var sw = Stopwatch.StartNew(); _byteArraysQueue=new ConcurrentQueue<byte[]>(); _keyQueue=new ConcurrentQueue<string>();
//生產key 和 消費key的兩個任務 var taskKeys = Task.Factory.StartNew(()=>ParallelGennerateMD5Keys(Environment.ProcessorCount - 1)); var taskHexString = Task.Factory.StartNew(()=>ConverKeysToHex(taskKeys)); string lastKey;
//隔半秒去看一次。 while (taskHexString.Status == TaskStatus.Running || taskHexString.Status == TaskStatus.WaitingToRun) { Console.WriteLine("_keyqueue的個數是{0},_byteArraysQueue的個數是{1}", _keyQueue.Count,_byteArraysQueue.Count); if (_keyQueue.TryPeek(out lastKey)) { // Console.WriteLine("第一個Key是{0}",lastKey); } Thread.Sleep(500); } //等待兩個任務結束 Task.WaitAll(taskKeys, taskHexString); Console.WriteLine("結束時間:" + sw.Elapsed); Console.WriteLine("key的總數是{0}" , _keyQueue.Count); Console.ReadKey(); }
從結果能夠發現,_bytaArraysQueue裏面的byte[] 幾乎是生產一個,就被消費一個。
理解生產者和消費者
使用ConcurrentQueue能夠很容易的實現並行的生產者-消費者模式或多階段的線性流水線。以下:
咱們能夠改造上面的main方法,讓一半的線程用於生產,一半的線程用於消費。
static void Main(string[] args) { Console.WriteLine("任務開始..."); var sw = Stopwatch.StartNew(); _byteArraysQueue=new ConcurrentQueue<byte[]>(); _keyQueue=new ConcurrentQueue<string>(); var taskKeyMax = Environment.ProcessorCount/2; var taskKeys = Task.Factory.StartNew(() => ParallelGennerateMD5Keys(taskKeyMax)); var taskHexMax = Environment.ProcessorCount - taskKeyMax; var taskHexStrings=new Task[taskHexMax]; for (int i = 0; i < taskHexMax; i++) { taskHexStrings[i] = Task.Factory.StartNew(() => ConverKeysToHex(taskKeys)); } Task.WaitAll(taskHexStrings); Console.WriteLine("結束時間:" + sw.Elapsed); Console.WriteLine("key的總數是{0}" , _keyQueue.Count); Console.ReadKey(); }
而這些消費者的結果又能夠繼續做爲生產者,繼續串聯下去。
ConcurrentStack
一個LIFO(後進先出)的集合,支持多任務併發進行壓入和彈出操做。它是徹底無鎖的。是System.Collections.Stack的併發版本。
它和ConcurrentQueue很是類似,區別在於使用了不一樣的方法名,更好的表示一個棧。ConcurrentStack主要提供了下面五個重要方法。
爲了判斷棧是否包含任意項,可使用IsEmpty屬性判斷。
if(!_byteArraysStack.IsEmpty)
而使用Count方法,開銷相對較大。另外咱們能夠將不安全的集合或數組轉化爲併發集合。下例將數組做爲參數傳入。操做上和List同樣。
private static string[] _HexValues = {"AF", "BD", "CF", "DF", "DA", "FE", "FF", "FA"}; static void Main(string[] args) { var invalidHexStack = new ConcurrentStack<string>(_HexValues); while (!invalidHexStack.IsEmpty) { string value; invalidHexStack.TryPop(out value); Console.WriteLine(value); } }
反之,能夠用CopyTo和ToArray方法將併發集合建立一個不安全集合。
ConcurrentBag
一個無序對象集合,在同一個線程添加元素(生產)和刪除元素(消費)的場合下效率特別高,ConcurrentBag最大程度上減小了同步的需求以及同步帶來的開銷。然而它在生產線程和消費線程徹底分開的狀況下,效率低下。
它提供了3個重要方法
下面的實例中Main方法經過Parallel.Invoke併發的加載三個方法。有多個生產者和消費者。對應三個ConcurrentBag<string>:_sentencesBag,_capWrodsInSentenceBag和_finalSentencesBag。
static void Main(string[] args) { Console.WriteLine("任務開始..."); var sw = Stopwatch.StartNew(); _sentencesBag=new ConcurrentBag<string>(); _capWrodsInSentenceBag=new ConcurrentBag<string>(); _finalSentencesBag=new ConcurrentBag<string>(); _producingSentences = true; Parallel.Invoke(ProduceSentences,CapitalizeWordsInSentence,RemoveLettersInSentence); Console.WriteLine("_sentencesBag的總數是{0}", _sentencesBag.Count); Console.WriteLine("_capWrodsInSentenceBag的總數是{0}", _capWrodsInSentenceBag.Count); Console.WriteLine("_finalSentencesBag的總數是{0}", _finalSentencesBag.Count); Console.WriteLine("總時間:{0}",sw.Elapsed); Console.ReadKey(); }
private static ConcurrentBag<string> _sentencesBag; private static ConcurrentBag<string> _capWrodsInSentenceBag; private static ConcurrentBag<string> _finalSentencesBag; private static volatile bool _producingSentences = false; private static volatile bool _capitalWords = false; private static void ProduceSentences() { string[] rawSentences = { "併發集合你可知", "ConcurrentBag 你值得擁有", "stoneniqiu", "博客園", ".Net併發編程學習", "Reading for you", "ConcurrentBag 是個無序集合" }; try { Console.WriteLine("ProduceSentences..."); _sentencesBag = new ConcurrentBag<string>(); var random = new Random(); for (int i = 0; i < NUM_AES_KEYS; i++) { var sb = new StringBuilder(); sb.Append(rawSentences[random.Next(rawSentences.Length)]); sb.Append(' '); _sentencesBag.Add(sb.ToString()); } } finally { _producingSentences = false; } } private static void CapitalizeWordsInSentence() { SpinWait.SpinUntil(() => _producingSentences); try { Console.WriteLine("CapitalizeWordsInSentence..."); _capitalWords = true; while ((!_sentencesBag.IsEmpty)||_producingSentences) { string sentence; if (_sentencesBag.TryTake(out sentence)) { _capWrodsInSentenceBag.Add(sentence.ToUpper()+"stoneniqiu"); } } } finally { _capitalWords = false; } } private static void RemoveLettersInSentence() { SpinWait.SpinUntil(() => _capitalWords); Console.WriteLine("RemoveLettersInSentence..."); while (!_capWrodsInSentenceBag.IsEmpty || _capitalWords) { string sentence; if (_capWrodsInSentenceBag.TryTake(out sentence)) { _finalSentencesBag.Add(sentence.Replace("stonenqiu","")); } } }
在CapitalizeWordsInSentence 方法中,使用SpinUntil方法並傳入共享bool變量_producingSentences,當其爲true的時候,SpinUnit方法會中止自旋。但協調多個生產者和消費者自旋並不是最好的解決方案,咱們可使用BlockingCollection(下面會講)來提高性能。
SpinWait.SpinUntil(() => _producingSentences);
另外兩個用做標誌的共享bool變量在聲明的時候使用了volatile關鍵字。這樣能夠確保在不一樣的線程中進行訪問的時候,能夠獲得這些變量的最新值。
private static volatile bool _producingSentences = false; private static volatile bool _capitalWords = false;
BlockingCollection
與經典的阻塞隊列數據結構相似,適用於多個任務添加和刪除數據的生產者-消費者的情形。提供了阻塞和界限的能力。
BlockingCollection是對IProducerConsumerCollection<T>實例的一個包裝。而這個接口繼承於ICollection,IEnumerable<T>。前面的併發集合都繼承了這個接口。所以這些集合均可以封裝在BlockingCollection中。
將上面的例子換成BlockingCollection
static void Main(string[] args) { Console.WriteLine("任務開始..."); var sw = Stopwatch.StartNew(); _sentencesBC = new BlockingCollection<string>(NUM_SENTENCE); _capWrodsInSentenceBC = new BlockingCollection<string>(NUM_SENTENCE); _finalSentencesBC = new BlockingCollection<string>(NUM_SENTENCE); Parallel.Invoke(ProduceSentences,CapitalizeWordsInSentence,RemoveLettersInSentence); Console.WriteLine("_sentencesBag的總數是{0}", _sentencesBC.Count); Console.WriteLine("_capWrodsInSentenceBag的總數是{0}", _capWrodsInSentenceBC.Count); Console.WriteLine("_finalSentencesBag的總數是{0}", _finalSentencesBC.Count); Console.WriteLine("總時間:{0}",sw.Elapsed); Console.ReadKey(); } private static int NUM_SENTENCE = 2000000; private static BlockingCollection<string> _sentencesBC; private static BlockingCollection<string> _capWrodsInSentenceBC; private static BlockingCollection<string> _finalSentencesBC; private static volatile bool _producingSentences = false; private static volatile bool _capitalWords = false; private static void ProduceSentences() { string[] rawSentences = { "併發集合你可知", "ConcurrentBag 你值得擁有", "stoneniqiu", "博客園", ".Net併發編程學習", "Reading for you", "ConcurrentBag 是個無序集合" }; Console.WriteLine("ProduceSentences..."); _sentencesBC = new BlockingCollection<string>(); var random = new Random(); for (int i = 0; i < NUM_SENTENCE; i++) { var sb = new StringBuilder(); sb.Append(rawSentences[random.Next(rawSentences.Length)]); sb.Append(' '); _sentencesBC.Add(sb.ToString()); } //讓消費者知道,生產過程已經完成 _sentencesBC.CompleteAdding(); } private static void CapitalizeWordsInSentence() { Console.WriteLine("CapitalizeWordsInSentence..."); //生產者是否完成 while (!_sentencesBC.IsCompleted) { string sentence; if (_sentencesBC.TryTake(out sentence)) { _capWrodsInSentenceBC.Add(sentence.ToUpper() + "stoneniqiu"); } } //讓消費者知道,生產過程已經完成 _capWrodsInSentenceBC.CompleteAdding(); } private static void RemoveLettersInSentence() { //SpinWait.SpinUntil(() => _capitalWords); Console.WriteLine("RemoveLettersInSentence..."); while (!_capWrodsInSentenceBC.IsCompleted) { string sentence; if (_capWrodsInSentenceBC.TryTake(out sentence)) { _finalSentencesBC.Add(sentence.Replace("stonenqiu","")); } } }
無需再使用共享的bool變量來同步。在操做結束後,調用CompeteAdding方法來告之下游的消費者。這個時候IsAddingComplete屬性爲true。
_sentencesBC.CompleteAdding();
而在生產者中也無需使用自旋了。能夠判斷IsCompleted屬性。而當IsAddingComplete屬性爲true且集合爲空的時候,IsCompleted才爲true。這個時候就表示,生產者的元素已經被使用完了。這樣代碼也更簡潔了。
while (!_sentencesBC.IsCompleted)
最後的結果要比使用ConcurrentBag快了0.8秒。一共是200w條數據,處理三次。
ConcurrentDictionary
與經典字典相似,提供了併發的鍵值訪問。它對讀操做是徹底無鎖的,在添加和修改的時候使用了細粒度的鎖。是IDictionary的併發版本。
它提供最重要方法以下:
下面的例子建立一個ConcurrentDictionary,而後不斷的更新。lock關鍵字確保一次只有一個線程運行Update方法。
static void Main(string[] args) { Console.WriteLine("任務開始..."); var sw = Stopwatch.StartNew(); rectangInfoDic=new ConcurrentDictionary<string, RectangInfo>(); GenerateRectangles(); foreach (var keyValue in rectangInfoDic) { Console.WriteLine("{0},{1},更新次數{2}",keyValue.Key,keyValue.Value.Size,keyValue.Value.UpdateTimes); } Console.WriteLine("總時間:{0}",sw.Elapsed); Console.ReadKey(); } private static ConcurrentDictionary<string, RectangInfo> rectangInfoDic; private const int MAX_RECTANGLES = 2000; private static void GenerateRectangles() { Parallel.For(1, MAX_RECTANGLES + 1, (i) => { for (int j = 0; j < 50; j++) { var newkey = string.Format("Rectangle{0}", i%5000); var rect = new RectangInfo(newkey, i, j); rectangInfoDic.AddOrUpdate(newkey, rect, (key, existRect) => { if (existRect != rect) { lock (existRect) { existRect.Update(rect.X,rect.Y); } return existRect; } return existRect; }); } }); }
Rectangle:
public class RectangInfo:IEqualityComparer<RectangInfo> { public string Name { get; set; } public int X { get; set; } public int Y { get; set; } public int UpdateTimes { get; set; } public int Size { get { return X*Y; } } public DateTime LastUpdate { get; set; } public RectangInfo(string name,int x,int y) { Name = name; X = x; Y = y; LastUpdate = DateTime.Now; } public RectangInfo(string key) : this(key, 0, 0) { } public void Update(int x,int y) { X = x; Y = y; UpdateTimes++; } public bool Equals(RectangInfo x, RectangInfo y) { return (x.Name == y.Name && x.Size == y.Size); } public int GetHashCode(RectangInfo obj) { return obj.Name.GetHashCode(); } }
本章學習了五種併發集合,熟悉了生產者-消費者的併發模型,咱們可使用併發集合來設計並優化流水線。但願本文對你有幫助。
閱讀書籍:《C#並行編程高級教程》 。