什么是MapReduce? Google的分布运算开发工具!

 

MapReduce是Google开发的C++编程工具，用于大规模数据集（大于1TB）的并行运算。我关注MapReduce已经很久了，前些日子开始翻译Wikipedia上面的介绍文章，但是由于忙于其他的事务，直到今天才彻底翻译完成，更新了 中文维基后，发在自己的Blog上，一方面多一个备份，另一方面方便不能访问维基的朋友查看，再有就是本人翻译水平和技术功底都不够，把 译文和 原文放在这里，有什么谬误请大家帮助更新维基上面的文章，不能访问维基的留言告知，我会尽快地更新，以免错误的理解和词语应用给大家带来误导。

MapReduce 中文版

MapReduce是Google开发的C++编程工具，用于大规模数据集（大于1TB）的并行运算。概念"Map（映射）"和"Reduce（化简）"，和他们的主要思想，都是从函数式编程语言里借来的，还有从矢量编程语言里借来的特性。[1]

当前的软件实现是指定一个Map（映射）函数，用来把一组键值对映射成一组新的键值对，指定并发的Reduce（化简）函数，用来保证所有映射的键值对中的每一个共享相同的键组。

映射和化简

简单说来，一个映射函数就是对一些独立元素组成的概念上的列表（例如，一个测试成绩的列表）的每一个元素进行指定的操作（比如前面的例子里，有人发现所有学生的成绩都被高估了一分，他可以定义一个“减一”的映射函数，用来修正这个错误。）。事实上，每个元素都是被独立操作的，而原始列表没有被更改，因为这里创建了一个新的列表来保存新的答案。这就是说，Map操作是可以高度并行的，这对高性能要求的应用以及并行计算领域的需求非常有用。

而化简操作指的是对一个列表的元素进行适当的合并（继续看前面的例子，如果有人想知道班级的平均分该怎么做？他可以定义一个化简函数，通过让列表中的元素跟自己的相邻的元素相加的方式把列表减半，如此递归运算直到列表只剩下一个元素，然后用这个元素除以人数，就得到了平均分。）。虽然他不如映射函数那么并行，但是因为化简总是有一个简单的答案，大规模的运算相对独立，所以化简函数在高度并行环境下也很有用。

分布和可靠性

MapReduce通过把对数据集的大规模操作分发给网络上的每个节点实现可靠性；每个节点会周期性的把完成的工作和状态的更新报告回来。如果一个节点保持沉默超过一个预设的时间间隔，主节点（类同Google File System中的主服务器）记录下这个节点状态为死亡，并把分配给这个节点的数据发到别的节点。每个操作使用命名文件的原子操作以确保不会发生并行线程间的冲突；当文件被改名的时候，系统可能会把他们复制到任务名以外的另一个名字上去。（避免副作用）。

化简操作工作方式很类似，但是由于化简操作在并行能力较差，主节点会尽量把化简操作调度在一个节点上，或者离需要操作的数据尽可能进的节点上了；这个特性可以满足Google的需求，因为他们有足够的带宽，他们的内部网络没有那么多的机器。

用途

在Google，MapReduce用在非常广泛的应用程序中，包括“分布grep，分布排序，web连接图反转，每台机器的词矢量，web访问日志分析，反向索引构建，文档聚类,机器学习，基于统计的机器翻译...”值得注意的是，MapReduce实现以后，它被用来重新生成Google的整个索引，并取代老的ad hoc程序去更新索引。

MapReduce会生成大量的临时文件，为了提高效率，它利用Google文件系统来管理和访问这些文件。

其他实现

Nutch项目开发了一个实验性的MapReduce的实现[2]。

参考

* Dean, Jeffrey & Ghemawat, Sanjay (2004)." MapReduce:大规模集群上的简单数据处理方式" 2005年4月6日。

^ "我们的灵感来自lisp和其他函数式编程语言中的古老的映射和化简操作." -"MapReduce:大规模集群上的简单数据处理方式"

MapReduce 英文版

MapReduce is a programming tool developed by Google in C++, in which parallel computations over large (> 1 terabyte) data sets are performed. The terminology of "Map" and "Reduce", and their general idea, is borrowed from functional programming languages use of the constructs map and reduce in functional programming and features of array programming languages. [1]

The actual software is implemented by specifying a Map function that maps key-value pairs to new key-value pairs and a subsequent Reduce function that consolidates all mapped key-value pairs sharing the same keys to single key-value pairs.

Map and Reduce

In simpler terms, what a map function does is go over a conceptual list of independent elements (for example, a list of test scores) and performs a specified operation on each element (with the previous example, one might have discovered a flaw in the test that gave each student a score too high by one; one could then define a map function of "minus 1"- it would subtract one from each score, correcting them.); the fact that each element is operated on independently, and that the original list is not being modified because a new list is created to hold the answers means that it is very easy to make a map operation highly parallel, and thus useful in high-performance applications and domains like parallel programming.

A reduce operation on the other hand, usually takes a list and combines elements appropriately (Continuing the preceding example, what if one wanted to know the class average? One could define a reduce function which halved the size of the list by adding an entry in the list to its neighbor, recursively continuing until there is only one (large) entry, and dividing the total sum by the original entry of elements to get the average); while since a reduce always ends up with a single answer, it is not as parallelizable as a map function, the large number of fairly independent calculations means that reduce functions are still useful in highly parallel environments.

Distribution and reliability

MapReduce achieves reliability by parceling out a number of operations on the set of data to each node in the network; each node is expected to report back periodically with completed work and status updates. If a node falls silent for longer than that interval, the master node (similar to the master server in the Google File System) records the node as dead, and sends out the node's assigned data to other nodes. Individual operations use atomic operations for naming file outputs as a double check to insure that there are not parallel conflicting threads running; when files are renamed, it is possible to also copy them to another name in addition to the name of the task (allowing for side-effects).

The reduce operations operate much the same way, but because of their inferior properties with regard to parallel operations, the master node attempts to schedule reduce operations on the same node, or as close as possible to the node holding the data being operated on; this property is desirable for Google as it conserves bandwidth, which their internal networks do not have much of.

Uses

According to Google, they use MapReduce in a wide range of applications, including: "distributed grep, distributed sort, web link-graph reversal, term-vector per host, web access log stats, inverted index construction, document clustering, machine learning, statistical machine translation..." Most significantly, when MapReduce was finished, it was used to completely regenerate Google's index of the Internet, and replaced the old ad hoc programs that updated the index.

MapReduce generates a large number of intermediate, temporary files, which are generally managed by, and accessed through, Google File System, for greater performance.

Other Implementations

The Nutch project has developed an experimental implementation [2] of MapReduce.
[edit]

References

* Dean, Jeffrey & Ghemawat, Sanjay (2004). "MapReduce: Simplified Data Processing on Large Clusters". Retrieved Apr. 6, 2005.

↑ "Our abstraction is inspired by the map and reduce primitives present in Lisp and many other functional languages." -"MapReduce: Simplified Data Processing on Large Clusters"

http://www.tinydust.net/prog/diary/2006/06/mapreduce-google.html

原文 http://blog.csdn.net/tinydust/archive/2006/06/11/787671.aspx

什么是MapReduce? Google的分布运算开发工具!