先看段代码:

1 for (int i = 0; i < 10; i++)
2 {
3 Task.Factory.StartNew(()=>Console.WriteLine($"{Thread.CurrentThread.ManagedThreadId} ~ {i}"));
4 }

从代码上可以看出我们预期是打印1~10,但实际的打印结果是:

 1 7 ~ 10
2 4 ~ 10
3 10 ~ 10
4 9 ~ 10
5 4 ~ 10
6 3 ~ 10
7 5 ~ 10
8 9 ~ 10
9 6 ~ 10
10 8 ~ 10

与预期的不一致,我们预期是打印数字1到10,但实际打印出来的是10次10。因为这几个lambda表达式中使用了同一个变量,并且这些匿名函数共享变量值。

再来看下面这段代码:

1 Action<int> displayNumber = n => Console.WriteLine(n);
2 int i = 5;
3 Task taskOne = Task.Factory.StartNew(() => displayNumber(i));
4 i = 7;
5 Task taskTwo = Task.Factory.StartNew(() => displayNumber(i));
6 Task.WaitAll(taskOne,taskTwo);

输出结果:

7
7

当闭包通过lambda表达式捕获可变变量时,lambda捕获变量的引用,而不是捕获该变量的当前值。因此,如果任务在变量的引用值更改后运行,则该值将是内存中最新的值,而不是捕获变量时的值。

为解决该问题,我们引入Parallel类来解决问题:

1 Parallel.For(0,10,i=>Console.WriteLine($"{Thread.CurrentThread.ManagedThreadId} ~ {i}"));

打印结果:

 1 1 ~ 0
2 1 ~ 2
3 3 ~ 1
4 3 ~ 4
5 3 ~ 7
6 3 ~ 8
7 3 ~ 9
8 1 ~ 3
9 5 ~ 5
10 4 ~ 6

Parallel 类 提供对并行循环和区域的支持, 现在我们看下Parallel.for的代码:

  1 // this needs to be in try-block because it can throw in  BuggyScheduler.MaxConcurrencyLevel
2 rootTask = new ParallelForReplicatingTask(
3 parallelOptions,
4 delegate
5 {
6 //
7 // first thing we do upon enterying the task is to register as a new "RangeWorker" with the
8 // shared RangeManager instance.
9 //
10 // If this call returns a RangeWorker struct which wraps the state needed by this task
11 //
12 // We need to call FindNewWork32() on it to see whether there's a chunk available.
13 //
14 // Cache some information about the current task
15 Task currentWorkerTask = Task.InternalCurrent;
16 bool bIsRootTask = (currentWorkerTask == rootTask);
17 RangeWorker currentWorker = new RangeWorker();
18 Object savedStateFromPreviousReplica = currentWorkerTask.SavedStateFromPreviousReplica;
19 if (savedStateFromPreviousReplica is RangeWorker)
20 currentWorker = (RangeWorker)savedStateFromPreviousReplica;
21 else
22 currentWorker = rangeManager.RegisterNewWorker();
23 // These are the local index values to be used in the sequential loop.
24 // Their values filled in by FindNewWork32
25 int nFromInclusiveLocal;
26 int nToExclusiveLocal;
27 if (currentWorker.FindNewWork32(out nFromInclusiveLocal, out nToExclusiveLocal) == false ||
28 sharedPStateFlags.ShouldExitLoop(nFromInclusiveLocal) == true)
29 {
30 return; // no need to run
31 }
32 // ETW event for ParallelFor Worker Fork
33 if (TplEtwProvider.Log.IsEnabled())
34 {
35 TplEtwProvider.Log.ParallelFork((currentWorkerTask != null ? currentWorkerTask.m_taskScheduler.Id : TaskScheduler.Current.Id), (currentWorkerTask != null ? currentWorkerTask.Id : 0),
36 forkJoinContextID);
37 }
38 TLocal localValue = default(TLocal);
39 bool bLocalValueInitialized = false; // Tracks whether localInit ran without exceptions, so that we can skip localFinally if it wasn't
40 try
41 {
42 // Create a new state object that references the shared "stopped" and "exceptional" flags
43 // If needed, it will contain a new instance of thread-local state by invoking the selector.
44 ParallelLoopState32 state = null;
45 if (bodyWithState != null)
46 {
47 Contract.Assert(sharedPStateFlags != null);
48 state = new ParallelLoopState32(sharedPStateFlags);
49 }
50 else if (bodyWithLocal != null)
51 {
52 Contract.Assert(sharedPStateFlags != null);
53 state = new ParallelLoopState32(sharedPStateFlags);
54 if (localInit != null)
55 {
56 localValue = localInit();
57 bLocalValueInitialized = true;
58 }
59 }
60 // initialize a loop timer which will help us decide whether we should exit early
61 LoopTimer loopTimer = new LoopTimer(rootTask.ActiveChildCount);
62 // Now perform the loop itself.
63 do
64 {
65 if (body != null)
66 {
67 for (int j = nFromInclusiveLocal;
68 j < nToExclusiveLocal && (sharedPStateFlags.LoopStateFlags == ParallelLoopStateFlags.PLS_NONE // fast path check as SEL() doesn't inline
69 || !sharedPStateFlags.ShouldExitLoop()); // the no-arg version is used since we have no state
70 j += 1)
71 {
72 body(j);
73 }
74 }
75 else if (bodyWithState != null)
76 {
77 for (int j = nFromInclusiveLocal;
78 j < nToExclusiveLocal && (sharedPStateFlags.LoopStateFlags == ParallelLoopStateFlags.PLS_NONE // fast path check as SEL() doesn't inline
79 || !sharedPStateFlags.ShouldExitLoop(j));
80 j += 1)
81 {
82 state.CurrentIteration = j;
83 bodyWithState(j, state);
84 }
85 }
86 else
87 {
88 for (int j = nFromInclusiveLocal;
89 j < nToExclusiveLocal && (sharedPStateFlags.LoopStateFlags == ParallelLoopStateFlags.PLS_NONE // fast path check as SEL() doesn't inline
90 || !sharedPStateFlags.ShouldExitLoop(j));
91 j += 1)
92 {
93 state.CurrentIteration = j;
94 localValue = bodyWithLocal(j, state, localValue);
95 }
96 }
97 // Cooperative multitasking hack for AppDomain fairness.
98 // Check if allowed loop time is exceeded, if so save current state and return. The self replicating task logic
99 // will detect this, and queue up a replacement task. Note that we don't do this on the root task.
100 if (!bIsRootTask && loopTimer.LimitExceeded())
101 {
102 currentWorkerTask.SavedStateForNextReplica = (object)currentWorker;
103 break;
104 }
105 }
106 // Exit if we can't find new work, or if the loop was stoppped.
107 while (currentWorker.FindNewWork32(out nFromInclusiveLocal, out nToExclusiveLocal) &&
108 ((sharedPStateFlags.LoopStateFlags == ParallelLoopStateFlags.PLS_NONE) ||
109 !sharedPStateFlags.ShouldExitLoop(nFromInclusiveLocal)));
110 }
111 catch
112 {
113 // if we catch an exception in a worker, we signal the other workers to exit the loop, and we rethrow
114 sharedPStateFlags.SetExceptional();
115 throw;
116 }
117 finally
118 {
119 // If a cleanup function was specified, call it. Otherwise, if the type is
120 // IDisposable, we will invoke Dispose on behalf of the user.
121 if (localFinally != null && bLocalValueInitialized)
122 {
123 localFinally(localValue);
124 }
125 // ETW event for ParallelFor Worker Join
126 if (TplEtwProvider.Log.IsEnabled())
127 {
128 TplEtwProvider.Log.ParallelJoin((currentWorkerTask != null ? currentWorkerTask.m_taskScheduler.Id : TaskScheduler.Current.Id), (currentWorkerTask != null ? currentWorkerTask.Id : 0),
129 forkJoinContextID);
130 }
131 }
132 },
133 creationOptions, internalOptions);
134 rootTask.RunSynchronously(parallelOptions.EffectiveTaskScheduler); // might throw TSE
135 rootTask.Wait();
136 // If we made a cancellation registration, we need to clean it up now before observing the OCE
137 // Otherwise we could be caught in the middle of a callback, and observe PLS_STOPPED, but oce = null
138 if (parallelOptions.CancellationToken.CanBeCanceled)
139 {
140 ctr.Dispose();
141 }
142 // If we got through that with no exceptions, and we were canceled, then
143 // throw our cancellation exception
144 if (oce != null) throw oce;

body对于迭代范围 (的每个值调用一次委托 fromInclusive , toExclusive) 。提供两个参数:

1、一个 Int32 值,该值表示迭代次数。

2、ParallelLoopState可用于提前中断循环的实例。ParallelLoopState对象是由编译器创建的; 它不能在用户代码中实例化。

继续来看:

Parallel.For(0, 10, (i,state) =>
{
if (i > 5)
state.Break();
Console.WriteLine($"{Thread.CurrentThread.ManagedThreadId} ~ {i}");
} );

输出:

1 1 ~ 0
2 1 ~ 1
3 1 ~ 2
4 1 ~ 3
5 1 ~ 4
6 1 ~ 5
7 1 ~ 6

在上面的方法中我们使用了 break方法。

调用 Break 方法会通知 for 操作,在当前的迭代之后,无需执行迭代。不过,如果所有迭代尚未执行,则仍必须执行当前的所有迭代。

因此,调用 Break 类似于 for c# 等语言中的传统循环内的中断操作,但它并不是完美的替代方法:例如,无法保证当前的迭代不会执行。

今天就先写道这里~


多线程那点事—Parallel.for的更多相关文章

  1. 关于JAVA多线程的那些事__初心者

    前言 其实事情的经过也许会复杂了点,这事还得从两个月前开始说.那天,我果断不干IT支援.那天,我立志要做一个真正的程序猿.那天,我26岁11个月.那天,我开始看Android.那天,我一边叨念着有朋自 ...

  2. 【QT】 Qt多线程的“那些事”

    目录 一.前言 二.QThread源码浅析 2.1 QThread类的定义源码 2.2 QThread::start()源码 2.3 QThreadPrivate::start()源码 2.4 QTh ...

  3. 【Python爬虫实战】多线程爬虫---糗事百科段子爬取

    多线程爬虫:即程序中的某些程序段并行执行,合理地设置多线程,可以让爬虫效率更高糗事百科段子普通爬虫和多线程爬虫分析该网址链接得出:https://www.qiushibaike.com/8hr/pag ...

  4. java多线程那点事

    屌丝程序员们对自己的技术能力总是毫不掩饰的高调,更有甚者每当完成一个简单的功能或算法实现,恨不得从工位上跳起来,生怕谁不知道一样,心情能理解,但个人完全鄙视这种行为.说到底,大家日常的coding,大 ...

  5. linux 多线程那点事

    说明:对多线程与相互排斥锁不熟悉的请參考其他 #include <pthread.h> #include <stdio.h> #include <stdlib.h> ...

  6. Java多线程编程那些事:volatile解惑--转

    http://www.infoq.com/cn/articles/java-multi-thread-volatile/ 1. 前言 volatile关键字可能是Java开发人员“熟悉而又陌生”的一个 ...

  7. C#多线程那点事——信号量(Semaphore)

    信号量说简单点就是为了线程同步,或者说是为了限制线程能运行的数量. 那它又是怎么限制线程的数量的哩?是因为它内部有个计数器,比如你想限制最多5个线程运行,那么这个计数器的值就会被设置成5,如果一个线程 ...

  8. .Net进阶系列(10)-异步多线程综述(被替换)

    一. 综述 经过两个多个周的整理,异步多线程章节终于整理完成,如下图所示,主要从基本概念.委托的异步调用.Thread多线程.ThreadPool多线程.Task.Parallel并行计算.async ...

  9. OpenMP多线程linux下的使用,简单化

    http://hi.baidu.com/diwulechao/item/bc6d865c411b813c32e0a932 http://www.cnblogs.com/yangyangcv/archi ...

随机推荐

  1. Python正则表达式书写容易碰到的陷阱:\W*和\W*?匹配过程遇到的问题

    老猿在分析<Python正则表达式\W+和\W*匹配过程的深入分析>中的问题时,想到一个问题,如果"re.split('(\W*)','Hello,world')"的处 ...

  2. 第14.4节 使用IE浏览器获取网站访问的http信息

    上节<第14.3节 使用google浏览器获取网站访问的http信息>中介绍了使用Google浏览器怎么获取网站访问的http相关报文信息,本节介绍IE浏览器中怎么获取相关信息.以上节为基 ...

  3. pandas 处理缺失值(连续值取平均,离散值fillna"<unk>")

    # 2.1处理缺失值,连续值用均值填充 continuous_fillna_number = [] for i in train_null_ix: if(i in continuous_ix): me ...

  4. es6 数组新增方法

    1.Array.from(): 这个函数的作用是将类似数组的对象转化为数组,比如DOM对象 let arrayLike = {      "0":"TangSir&quo ...

  5. Win10新增API

    Win10新增API为以下一些方面: Windows.AI Windows.ApplicationModel Windows.Devices Windows.Globalization Windows ...

  6. SSM框架中常用的注解及含义

    @Controller---使用它标记在一个类上,dispatcher会扫描使用该注解类的方法,并检测该方法是否使用了@RequestMapping注解,加上RequestMapping注解的方法才是 ...

  7. Pytorch系列之常用基础操作

    各种张量初始化 创建特殊类型的tensor a = torch.FloatTensor(2,3) a = torch.DoubleTensor(2,3) ... 设置pytorch中tensor的默认 ...

  8. JavaSE01-概述

    1.1 Java语言发展史 语言:人与人交流沟通的表达方式 计算机语言:人与计算机之间进行信息交流沟通的一种特殊语言 Java语言是美国Sun公司(Stanford University Networ ...

  9. 持续引领大数据行业发展,腾讯云发布全链路数据开发平台WeData

    9月11日,在腾讯全球数字生态大会大数据专场上,腾讯云大数据产品副总经理雷小平重磅发布了全链路数据开发平台WeData,同时发布和升级了流计算服务.云数据仓库.ES.企业画像等6款核心产品,进一步优化 ...

  10. Graphql Tutorials(Episode 02)

    1.前言 我们在上篇已经了解Graphql的使命以及Graphql的概况,接下来,我们跑起来另外一个Helloworld来开启继续学习. 2.Helloworld(使用Graphql 原生API) 这 ...