Content deleted Content added
Michaelmalak (talk | contribs) Undid revision 1223249519 by JohnPritchard (talk) cite does not appear to be about map/reduce even though it contains both those words in the title |
|||
| (45 intermediate revisions by 31 users not shown) | |||
Line 1:
{{Short description|Parallel programming model}}
'''MapReduce''' is a [[programming model]] and an associated implementation for processing and generating [[big data]] sets with a [[Parallel computing|parallel]], [[distributed computing|distributed]] algorithm on a [[Cluster (computing)|cluster]].<ref>{{cite web|url=https://hadoop.apache.org/docs/r1.2.1/mapred_tutorial.html|title=MapReduce Tutorial|access-date=3 July 2019|website=Apache Hadoop}}</ref><ref>{{cite web|url=http://news.cnet.com/8301-10784_3-9955184-7.html|title=Google spotlights data center inner workings|date=30 May 2008|website=cnet.com|access-date=31 May 2008|archive-date=19 October 2013|archive-url=https://web.archive.org/web/20131019063218/http://news.cnet.com/8301-10784_3-9955184-7.html|url-status=dead}}</ref><ref name="GoogleMapReduce">{{cite web|url=http://static.googleusercontent.com/media/research.google.com/es/us/archive/mapreduce-osdi04.pdf|title=MapReduce: Simplified Data Processing on Large Clusters|website=googleusercontent.com}}</ref>
A MapReduce program is composed of a [[map (parallel pattern)|''map'']] [[procedure (computing)|procedure]], which performs filtering and sorting (such as sorting students by first name into queues, one queue for each name), and a ''[[Reduce (parallel pattern)|reduce]]'' method, which performs a summary operation (such as counting the number of students in each queue, yielding name frequencies). The "MapReduce System" (also called "infrastructure" or "framework") orchestrates the processing by [[Marshalling (computer science)|marshalling]] the distributed servers, running the various tasks in parallel, managing all communications and data transfers between the various parts of the system, and providing for [[Redundancy (engineering)|redundancy]] and [[Fault-tolerant computer system|fault tolerance]].
The model is a specialization of the ''split-apply-combine'' strategy for data analysis.<ref>{{Cite journal | doi = 10.18637/jss.v040.i01| title = The split-apply-combine strategy for data analysis| journal = Journal of Statistical Software| volume = 40| pages = 1–29| year = 2011| last1 = Wickham| first1 = Hadley | doi-access = free}}</ref>
It is inspired by the [[map (higher-order function)|map]] and [[reduce (higher-order function)|reduce]] functions commonly used in [[functional programming]],<ref name="map">"Our abstraction is inspired by the map and reduce primitives present in Lisp and many other functional languages." -[http://research.google.com/archive/mapreduce.html "MapReduce: Simplified Data Processing on Large Clusters"], by Jeffrey Dean and Sanjay Ghemawat; from Google Research</ref> although their purpose in the MapReduce framework is not the same as in their original forms.<ref>{{Cite journal | doi = 10.1016/j.scico.2007.07.001| title = Google's Map ''Reduce'' programming model — Revisited| journal = Science of Computer Programming| volume = 70| pages = 1–30| year = 2008| last1 = Lämmel | first1 = R. | doi-access =
| url = https://stackoverflow.com/questions/3947889/mongodb-terrible-mapreduce-performance
| title = MongoDB: Terrible MapReduce Performance
Line 24 ⟶ 25:
# '''Reduce:''' worker nodes now process each group of output data, per key, in parallel.
MapReduce allows for the distributed processing of the map and reduction operations. Maps can be performed in parallel, provided that each mapping operation is independent of the others; in practice, this is limited by the number of independent data sources and/or the number of CPUs near each source. Similarly, a set of 'reducers' can perform the reduction phase, provided that all outputs of the map operation that share the same key are presented to the same reducer at the same time, or that the reduction function is [[Associative property|associative]]. While this process often appears inefficient compared to algorithms that are more sequential (because multiple instances of the reduction process must be run), MapReduce can be applied to significantly larger datasets than a single [[Commodity computing|"commodity" server]] can handle – a large [[server farm]] can use MapReduce to sort a [[petabyte]] of data in only a few hours.<ref>{{cite web|last=Czajkowski|first=Grzegorz|title=Sorting Petabytes with MapReduce – The Next Episode|url=http://googleresearch.blogspot.com/2011/09/sorting-petabytes-with-mapreduce-next.html|access-date=7 April 2014|author2=Marián Dvorský |author3=Jerry Zhao |author4=Michael Conley |date=7 September 2011 }}</ref> The parallelism also offers some possibility of recovering from partial failure of servers or storage during the operation: if one mapper or reducer fails, the work can be rescheduled – assuming the input data are still available.
Another way to look at MapReduce is as a 5-step parallel and distributed computation:
Line 48 ⟶ 49:
The ''Reduce'' function is then applied in parallel to each group, which in turn produces a collection of values in the same domain:
<code>Reduce(k2, list (v2))</code> → <code>list((k3, v3))</code><ref>{{Cite web|url=https://hadoop.apache.org/docs/r1.2.1/mapred_tutorial.html#Inputs+and+Outputs|title = MapReduce Tutorial}}</ref>
Each ''Reduce'' call typically produces either one key value pair or an empty return, though one call is allowed to return more than one key value pair. The returns of all calls are collected as the desired result list.
Thus the MapReduce framework transforms a list of (key, value) pairs into another list of (key, value) pairs.<ref>{{Cite web|url=https://github.com/apache/hadoop-mapreduce/blob/307cb5b316e10defdbbc228d8cdcdb627191ea15/src/java/org/apache/hadoop/mapreduce/Reducer.java#L148|title = Apache/Hadoop-mapreduce|website = [[GitHub]]|date = 31 August 2021}}</ref> This behavior is different from the typical functional programming map and reduce combination, which accepts a list of arbitrary values and returns one single value that combines ''all'' the values returned by map.
It is [[Necessity and sufficiency|necessary but not sufficient]] to have implementations of the map and reduce abstractions in order to implement MapReduce. Distributed implementations of MapReduce require a means of connecting the processes performing the Map and Reduce phases. This may be a [[distributed file system]]. Other options are possible, such as direct streaming from mappers to reducers, or for the mapping processors to serve up their results to reducers that query them.
Line 58 ⟶ 59:
===Examples===
The canonical MapReduce example counts the appearance of each word in a set of documents:<ref>{{cite web|url=http://research.google.com/archive/mapreduce-osdi04-slides/index-auto-0004.html|title=Example: Count word occurrences|publisher=Google Research|access-date=September 18, 2013}}</ref>
'''function''' <u>map</u>(String name, String document):
''// name: document name''
''// document: document contents''
'''for each''' word w '''in''' document:
emit (w, 1)
'''function''' <u>reduce</u>(String word, Iterator partialCounts):
''// word: a word''
''// partialCounts: a list of aggregated partial counts''
sum = 0
'''for each''' pc '''in''' partialCounts:
sum += pc
emit (word, sum)
Here, each document is split into words, and each word is counted by the ''map'' function, using the word as the result key. The framework puts together all the pairs with the same key and feeds them to the same call to ''reduce''. Thus, this function just needs to sum all of its input values to find the total appearances of that word.
Line 85:
</syntaxhighlight>
Using MapReduce, the
'''function''' Map '''is'''
Line 106:
'''end function'''
Note that in the {{mono|Reduce}} function, {{mono|C}} is the count of people having in total N contacts, so in the {{mono|Map}} function it is natural to write {{mono|1=C=1}}, since every output pair is referring to the contacts of one single person.
The MapReduce system would line up the 1100 Map processors, and would provide each with its corresponding 1 million input records. The Map step would produce 1.1 billion <tt>(Y,(N,1))</tt> records, with <tt>Y</tt> values ranging between, say, 8 and 103. The MapReduce System would then line up the 96 Reduce processors by performing shuffling operation of the key/value pairs due to the fact that we need average per age, and provide each with its millions of corresponding input records. The Reduce step would result in the much reduced set of only 96 output records <tt>(Y,A)</tt>, which would be put in the final result file, sorted by <tt>Y</tt>.▼
▲The MapReduce system would line up the 1100 Map processors, and would provide each with its corresponding 1 million input records. The Map step would produce 1.1 billion
The count info in the record is important if the processing is reduced more than one time. If we did not add the count of the records, the computed average would be wrong, for example:
Line 122 ⟶ 124:
10, 10
If we reduce files
''-- reduce step #1: age, average of contacts''
10, 9
If we reduce it with file
==Dataflow==
Line 166 ⟶ 168:
===Output writer===
The ''Output Writer'' writes the output of the ''Reduce'' to the stable storage.
==Theoretical background==
Properties of [[Monoid|monoids]] are the basis for ensuring the validity of MapReduce operations.<ref>{{Cite journal
| doi = 10.1017/S0956796817000193
| title = An algebra for distributed Big Data analytics
| journal = Journal of Functional Programming
| volume = 28
| year = 2017
| last = Fegaras
| first = Leonidas
| s2cid = 44629767
| doi-access =
}}</ref><ref>{{cite arXiv
|last=Lin
|first=Jimmy
|title=Monoidify! Monoids as a Design Principle for Efficient MapReduce Algorithms
|eprint=1304.7544
|date=29 Apr 2013|class=cs.DC
}}</ref>
In the Algebird package<ref>{{Cite web
|title= Abstract Algebra for Scala
|url=https://twitter.github.io/algebird/}}</ref> a Scala implementation of Map/Reduce explicitly requires a monoid class type
.<ref>{{Cite web
|title= Encoding Map-Reduce As A Monoid With Left Folding
|date= 5 September 2016|url= http://erikerlandson.github.io/blog/2016/09/05/expressing-map-reduce-as-a-left-folding-monoid/}}</ref>
The operations of MapReduce deal with two types: the type ''A'' of input data being mapped, and the type ''B'' of output data being reduced.
The ''Map'' operation takes individual values of type ''A'' and produces, for each ''a:A'' a value ''b:B''; The ''Reduce'' operation requires a binary operation • defined on values of type ''B''; it consists of folding all available ''b:B'' to a single value.
From a basic requirements point of view, any MapReduce operation must involve the ability to arbitrarily regroup data being reduced. Such a requirement amounts to two properties of the operation •:
* associativity: (''x'' • ''y'') • ''z'' = ''x'' • (''y'' • ''z'')
* existence of neutral element ''e'' such that ''e'' • ''x'' = ''x'' • ''e'' = ''x'' for every ''x:B''.
The second property guarantees that, when parallelized over multiple nodes, the nodes that don't have any data to process would have no impact on the result.
These two properties amount to having a [[Monoid|monoid]] (''B'', •, ''e'') on values of type ''B'' with operation • and with neutral element ''e''.
There's no requirements on the values of type ''A''; an arbitrary function ''A'' → ''B'' can be used for the ''Map'' operation. This means that we have a [[Catamorphism|catamorphism]] ''A*'' → (''B'', •, ''e''). Here ''A*'' denotes a [[Kleene star]], also known as the type of lists over ''A''.
The ''Shuffle'' operation per se is not related to the essence of MapReduce; it's needed to distribute calculations over the cloud.
It follows from the above that not every binary ''Reduce'' operation will work in MapReduce. Here are the counter-examples:
* building a tree from subtrees: this operation is not associative, and the result will depend on grouping;
* direct calculation of averages: ''avg'' is also not associative (and it has no neutral element); to calculate an average, one needs to calculate [[Moment (mathematics)|moments]].
==Performance considerations==
MapReduce programs are not guaranteed to be fast. The main benefit of this programming model is to exploit the optimized shuffle operation of the platform, and only having to write the ''Map'' and ''Reduce'' parts of the program.
In practice, the author of a MapReduce program however has to take the shuffle step into consideration; in particular the partition function and the amount of data written by the ''Map'' function can have a large impact on the performance and scalability. Additional modules such as the ''Combiner'' function can help to reduce the amount of data written to disk, and transmitted over the network. MapReduce applications can achieve sub-linear speedups under specific circumstances.<ref name=":0">{{Cite journal|title = BSP cost and scalability analysis for MapReduce operations|journal = Concurrency and Computation: Practice and Experience|date = 2015-01-01|issn = 1532-0634|pages = 2503–2527|doi = 10.1002/cpe.3628|
When designing a MapReduce algorithm, the author needs to choose a good tradeoff<ref name="ullman">{{Cite journal | doi = 10.1145/2331042.2331053 | title = Designing good MapReduce algorithms| journal = XRDS: Crossroads, the ACM Magazine for Students| volume = 19| pages =
In tuning performance of MapReduce, the complexity of mapping, shuffle, sorting (grouping by the key), and reducing has to be taken into account. The amount of data produced by the mappers is a key parameter that shifts the bulk of the computation cost between mapping and reducing. Reducing includes sorting (grouping of the keys) which has nonlinear complexity. Hence, small partition sizes reduce sorting time, but there is a trade-off because having a large number of reducers may be impractical. The influence of split unit size is marginal (unless chosen particularly badly, say <1MB). The gains from some mappers reading load from local disks, on average, is minor.<ref>{{Cite journal|title = Scheduling divisible MapReduce computations|
For processes that complete quickly, and where the data fits into main memory of a single machine or a small cluster, using a MapReduce framework usually is not effective. Since these frameworks are designed to recover from the loss of whole nodes during the computation, they write interim results to distributed storage. This crash recovery is expensive, and only pays off when the computation involves many computers and a long runtime of the computation. A task that completes in seconds can just be restarted in the case of an error, and the likelihood of at least one machine failing grows quickly with the cluster size. On such problems, implementations keeping all data in memory and simply restarting a computation on node failures or —when the data is small enough— non-distributed solutions will often be faster than a MapReduce system.
Line 185 ⟶ 235:
==Uses==
MapReduce is useful in a wide range of applications, including distributed pattern-based searching, distributed sorting, web link-graph reversal, Singular Value Decomposition,<ref>{{cite
multi-cluster,<ref name="HMR">{{Cite book | doi = 10.1145/1996023.1996026| chapter = A Hierarchical Framework for Cross-Domain MapReduce Execution|chapter-url = http://yuanluo.net/publications/LUO_ECMLS2011.pdf| title = Proceedings of the second international workshop on Emerging computational methods for the life sciences (ECMLS '11)| year = 2011| last1 = Luo | first1 = Y. | last2 = Guo | first2 = Z. | last3 = Sun | first3 = Y.| last4 = Plale | first4 = B. |author4-link=Beth Plale| last5 = Qiu | first5 = J. | last6=Li|
first6=W. |isbn = 978-1-4503-0702-4| citeseerx = 10.1.1.364.9898| s2cid = 15179363}}</ref> volunteer computing environments,<ref name="volunteerMR">{{Cite book | doi = 10.1145/1851476.1851489| chapter = MOON: MapReduce On Opportunistic eNvironments| title = Proceedings of the 19th ACM International Symposium on High Performance Distributed Computing – HPDC '10| pages = 95| year = 2010| last1 = Lin | first1 = H. | last2 = Ma | first2 = X. | last3 = Archuleta | first3 = J. | last4 = Feng | first4 = W. C. | last5 = Gardner | first5 = M. | last6 = Zhang | first6 = Z. | chapter-url = http://eprints.cs.vt.edu/archive/00001089/01/moon.pdf| isbn = 9781605589428| s2cid = 2351790}}</ref> dynamic cloud environments,<ref name="dynCloudMR">{{Cite journal| doi = 10.1016/j.jcss.2011.12.021| title = P2P-MapReduce: Parallel data processing in dynamic Cloud environments| journal = [[Journal of Computer and System Sciences]]| volume = 78| issue = 5| pages = 1382–1402| year = 2012
At Google, MapReduce was used to completely regenerate Google's index of the [[World Wide Web]]. It replaced the old ''ad hoc'' programs that updated the index and ran the various analyses.<ref name="usage">{{cite web| quote=As of October, Google was running about 3,000 computing jobs per day through MapReduce, representing thousands of machine-days, according to a presentation by Dean. Among other things, these batch routines analyze the latest Web pages and update Google's indexes.| url=http://www.baselinemag.com/c/a/Infrastructure/How-Google-Works-1/5| title=How Google Works| date=7 July 2006| publisher=baselinemag.com}}</ref> Development at Google has since moved on to technologies such as Percolator, FlumeJava<ref name="Chambers2010">{{cite book |last1=Chambers |first1=Craig |last2=Raniwala |first2=Ashish |last3=Perry |first3=Frances |last4=Adams |first4=Stephen |last5=Henry |first5=Robert R. |last6=Bradshaw |first6=Robert |last7=Weizenbaum |first7=Nathan |title
MapReduce's stable inputs and outputs are usually stored in a [[distributed file system]]. The transient data are usually stored on local disk and fetched remotely by the reducers.
Line 202 ⟶ 252:
DeWitt and Stonebraker have subsequently published a detailed benchmark study in 2009 comparing performance of [[Hadoop|Hadoop's]] MapReduce and [[RDBMS]] approaches on several specific problems.<ref name="sigmod">{{cite web| url=http://database.cs.brown.edu/projects/mapreduce-vs-dbms| title=A Comparison of Approaches to Large-Scale Data Analysis|first1=Andrew |last1=Pavlo |first2=Erik |last2=Paulson |first3=Alexander |last3=Rasin |first4=Daniel J. |last4=Abadi |first5=Deavid J. |last5=DeWitt |first6=Samuel |last6=Madden |first7=Michael |last7=Stonebraker| publisher=Brown University| access-date=2010-01-11}}</ref> They concluded that relational databases offer real advantages for many kinds of data use, especially on complex processing or where the data is used across an enterprise, but that MapReduce may be easier for users to adopt for simple or one-time processing tasks.
The MapReduce programming paradigm was also described in [[Danny Hillis]]'s 1985 thesis <ref name="WDHmit86">{{cite book |author-first=W. Danny |author-last=Hillis |date=1986 |title=The Connection Machine |publisher=[[MIT Press]] |isbn=0262081571 |url-access=registration |url=https://archive.org/details/connectionmachin00hill }}</ref> intended for use on the [[Connection Machine]], where it was called "xapping/reduction"<ref>{{cite web |url=http://bitsavers.trailing-edge.com/pdf/thinkingMachines/CM2/HA87-4_Connection_Machine_Model_CM-2_Technical_Summary_Apr1987.pdf |title=Connection Machine Model CM-2 Technical Summary |author=<!--Not stated--> |date=1987-04-01 |publisher=[[Thinking Machines Corporation]] |access-date=2022-11-21}}</ref> and relied upon that machine's special hardware to accelerate both map and reduce. The dialect ultimately used for the Connection Machine, the 1986 [[StarLisp]], had parallel <code>*map</code> and <code>reduce!!</code>,<ref>{{cite web |url=https://www.softwarepreservation.org/projects/LISP/starlisp/supplement-to-the-starlisp-reference-manual-version-5-0.pdf |title=Supplement to the *Lisp Reference Manual |author=<!--Not stated--> |date=1988-09-01 |publisher=[[Thinking Machines Corporation]] |access-date=2022-11-21}}</ref> which in turn was based on the 1984 [[Common Lisp]], which had non-parallel <code>map</code> and <code>reduce</code> built in.<ref>{{cite web |url=https://collections.lib.utah.edu/dl_files/20/2e/202ebf04b52d043c78297444bc9bc4fbc17b6b5e.pdf |title=Rediflow Architecture Prospectus |author=<!--Not stated--> |date=1986-04-05 |publisher=[[University of Utah School of Computing|University of Utah Department of Computer Science]] |access-date=2022-11-21}}</ref> The [[Fold (higher-order function)#Linear vs. tree-like folds|tree-like]] approach that the Connection Machine's [[Hypercube internetwork topology|hypercube architecture]] uses to execute <code>reduce</code> in <math>O(\log n)</math> time<ref>{{cite book |url=https://www.cise.ufl.edu/~sahni/papers/imagemono.pdf#page=20 |title=Hypercube Algorithms for Image Processing and Pattern Recognition |last=Ranka |first=Sanjay |date=1989 |access-date=2022-12-08 |section=2.6 Data Sum |publisher=University of Florida}}</ref> is effectively the same as the approach referred to within the Google paper as prior work.{{r|GoogleMapReduce|p=11|q=an associative function can be computed over all prefixes of an N element array in log N time on N processors using parallel prefix computations. MapReduce can be considered a simplification and distillation of some of these models}}
===Restricted programming framework===
MapReduce tasks must be written as acyclic dataflow programs, i.e. a stateless mapper followed by a stateless reducer, that are executed by a batch job scheduler. This paradigm makes repeated querying of datasets difficult and imposes limitations that are felt in fields such as [[
==See also==
* [[Bird–Meertens formalism]]
* [[Parallelization contract]]
===Implementations of MapReduce===
Line 222 ⟶ 273:
{{reflist|30em}}
{{Commons category|MapReduce}}
{{Google
{{Authority control}}
| |||