MapReduce近几年比较热的分布式计算编程模型,以C#为例简单介绍下MapReduce分布式计算。
阅读目录
背景
某平行世界程序猿小张接到Boss一项任务,统计用户反馈内容中的单词出现次数,以便分析用户主要习惯。文本如下:
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
const string hamlet = @"Though yet of Hamlet our dear brother's death The memory be green, and that it us befitted To bear our hearts in grief and our whole kingdom To be contracted in one brow of woe, Yet so far hath discretion fought with nature That we with wisest sorrow think on him, Together with remembrance of ourselves. Therefore our sometime sister, now our queen, The imperial jointress to this warlike state, Have we, as 'twere with a defeated joy,-- With an auspicious and a dropping eye, With mirth in funeral and with dirge in marriage, In equal scale weighing delight and dole,-- Taken to wife: nor have we herein barr'd Your better wisdoms, which have freely gone With this affair along. For all, our thanks. Now follows, that you know, young Fortinbras, Holding a weak supposal of our worth, Or thinking by our late dear brother's death Our state to be disjoint and out of frame, Colleagued with the dream of his advantage, He hath not fail'd to pester us with message, Importing the surrender of those lands Lost by his father, with all bonds of law, To our most valiant brother. So much for him. Now for ourself and for this time of meeting: Thus much the business is: we have here writ To Norway, uncle of young Fortinbras,-- Who, impotent and bed-rid, scarcely hears Of this his nephew's purpose,--to suppress His further gait herein; in that the levies, The lists and full proportions, are all made Out of his subject: and we here dispatch You, good Cornelius, and you, Voltimand, For bearers of this greeting to old Norway; Giving to you no further personal power To business with the king, more than the scope Of these delated articles allow. Farewell, and let your haste commend your duty.";
小张作为蓝翔高材生,很快就实现了:
var content = hamlet.Split(new[] { " ", Environment.NewLine }, StringSplitOptions.RemoveEmptyEntries); var wordcount=new Dictionary<string,int>(); foreach (var item in content) { if (wordcount.ContainsKey(item)) wordcount[item] += 1; else wordcount.Add(item, 1); }
作为有上进心的青年,小张决心对算法进行抽象封装,并支持多节点计算。小张把这个统计次数程序分成两个大步骤:分解和计算。
第一步:先把文本以某维度分解映射成最小独立单元。 (段落、单词、字母维度)。
第二部:把最小单元重复的做合并计算。
小张参考MapReduce论文设计Map、Reduce如下:
Map实现
Mapping
Mapping函数把文本分解映射key,value形式的最小单元,即<单词,出现次数(1)>、<word,1>。
public IEnumerable<Tuple<T, int>> Mapping(IEnumerable<T> list) { foreach (T sourceVal in list) yield return Tuple.Create(sourceVal, 1); }
使用,输出为(brow, 1), (brow, 1), (sorrow, 1), (sorrow, 1):
var spit = hamlet.Split(new[] { " ", Environment.NewLine }, StringSplitOptions.RemoveEmptyEntries); var mp = new MicroMapReduce<string>(new Master<string>()); var result= mp.Mapping(spit);
Combine
为了减少数据通信开销,mapping出的键值对数据在进入真正的reduce前,进行重复键合并。也相对于提前进行预计算一部分,加快总体计算速度。 输出格式为(brow, 2), (sorrow, 2):
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public Dictionary<T, int> Combine(IEnumerable<Tuple<T, int>> list) { Dictionary<T, int> dt = new Dictionary<T, int>(); foreach (var val in list) { if (dt.ContainsKey(val.Item1)) dt[val.Item1] += val.Item2; else dt.Add(val.Item1, val.Item2); } return dt; }
Partitioner
Partitioner主要用来分组划分,把不同节点的统计数据按照key进行分组。
其输出格式为: (brow, {(brow,2)},(brow,3)), (sorrow, {(sorrow,10)},(brow,11)):
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public IEnumerable<Group<T, int>> Partitioner(Dictionary<T, int> list) { var dict = new Dictionary<T, Group<T, int>>(); foreach (var val in list) { if (!dict.ContainsKey(val.Key)) dict[val.Key] = new Group<T, int>(val.Key); dict[val.Key].Values.Add(val.Value); } return dict.Values; }
Group定义:
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public class Group<TKey, TValue> : Tuple<TKey, List<TValue>> { public Group(TKey key) : base(key, new List<TValue>()) { } public TKey Key { get { return base.Item1; } } public List<TValue> Values { get { return base.Item2; } } }
Reduce实现
Reducing函数接收,分组后的数据进行最后的统计计算。
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public Dictionary<T, int> Reducing(IEnumerable<Group<T, int>> groups) { Dictionary<T, int> result=new Dictionary<T, int>(); foreach (var sourceVal in groups) { result.Add(sourceVal.Key, sourceVal.Values.Sum()); } return result; }
封装调用如下:
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public IEnumerable<Group<T, int>> Map(IEnumerable<T> list) { var step1 = Mapping(list); var step2 = Combine(step1); var step3 = Partitioner(step2); return step3; } public Dictionary<T, int> Reduce(IEnumerable<Group<T, int>> groups) { var step1 = Reducing(groups); return step1; }
public Dictionary<T, int> MapReduce(IEnumerable<T> list) { var map = Map(list); var reduce = Reduce(map); return reduce; }
整体计算步骤图如下:
支持分布式
小张抽象封装后,虽然复杂度上去了。但暴露给使用者是非常清晰的接口,满足MapReduce的数据格式要求,即可使用。
var spit = hamlet.Split(new[] { " ", Environment.NewLine }, StringSplitOptions.RemoveEmptyEntries); var mp = new MicroMapReduce<string>(new Master<string>()); var result1= mp.MapReduce(spit);
小张完成后脑洞大开,考虑到以后文本数据量超大。 所以fork了个分支,准备支持分布式计算,以后可以在多个服务器节点跑。
数据分片
数据分片就是把大量数据拆成一块一块的,分散到各个节点上,方便我们的mapReduce程序去计算。 分片主流的有mod、consistent hashing、vitual Buckets、Range Partition等方式。 关于consistent hashing上篇有介绍(探索c#之一致性Hash详解)。在Hadoop中Hdfs和mapreduce是相互关联配合的,一个存储和一个计算。如果自行实现的话还需要个统一的存储。所以这里的数据源可以是数据库也可以是文件。小张只是满足boss需求,通用计算框架的话可以直接用现成的。
模拟分片
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public List<IEnumerable<T>> Partition(IEnumerable<T> list) { var temp =new List<IEnumerable<T>>(); temp.Add(list); temp.Add(list); return temp; }
Worker节点
小张定义了Master,worker角色。 master负责汇集输出,即我们的主程序。 每一个worker我们用一个线程来模拟,最后输出到master汇总,master最后可以写到数据库或其他。
public void WorkerNode(IEnumerable<T> list) { new Thread(() => { var map = Map(list); var reduce = Reduce(map); master.Merge(reduce); }).Start(); }
![](https://i-blog.csdnimg.cn/blog_migrate/8f900a89c6347c561fdf2122f13be562.gif)
![](https://i-blog.csdnimg.cn/blog_migrate/961ddebeb323a10fe0623af514929fc1.gif)
public class Master<T> { public Dictionary<T, int> Result = new Dictionary<T, int>(); public void Merge(Dictionary<T, int> list) { foreach (var item in list) { lock (this) { if (Result.ContainsKey(item.Key)) Result[item.Key] += item.Value; else Result.Add(item.Key, item.Value); } } } }
分布式计算步骤图:
总结
MapReduce模型从性能速度来说并不是非常好的,它优势在于隐藏了分布式计算的细节、容灾错误、负载均衡及良好的编程API,包含HDFS、Hive等在内一整套大数据处理的生态框架体系。在数据量级不是很大的话,企业自行实现一套轻量级分布式计算会有很多优点,比如性能更好、可定制化、数据库也不需要导入导出。从成本上也节省不少,因为hadoop开发、运维、服务器都需要不少人力物力。