《大规模数据处理——云计算 Mass Data Processing Cloud Computing》课程教学资源（阅读材料）MapReduce——Simplified Data Processing on Large Clusters

MapReduce:Simplified Data Processing on Large Clusters Jeffrey Dean and Sanjay Ghemawat jeff@google.com,sanjay@google.com Google,Inc. Abstract given day,etc.Most such computations are conceptu- ally straightforward.However,the input data is usually MapReduce is a programming model and an associ- large and the computations have to be distributed across ated implementation for processing and generating large hundreds or thousands of machines in order to finish in data sets.Users specify a map function that processes a a reasonable amount of time.The issues of how to par- key/value pair to generate a set of intermediate key/value allelize the computation,distribute the data,and handle pairs,and a reduce function that merges all intermediate failures conspire to obscure the original simple compu- values associated with the same intermediate key.Many tation with large amounts of complex code to deal with real world tasks are expressible in this model,as shown these issues in the paper. As a reaction to this complexity.we designed a new Programs written in this functional style are automati- abstraction that allows us to express the simple computa- cally parallelized and executed on a large cluster of com- tions we were trying to perform but hides the messy de- modity machines.The run-time system takes care of the tails of parallelization,fault-tolerance,data distribution details of partitioning the input data,scheduling the pro- and load balancing in a library.Our abstraction is in- gram's execution across a set of machines,handling ma- spired by the map and reduce primitives present in Lisp chine failures,and managing the required inter-machine and many other functional languages.We realized that communication.This allows programmers without any most of our computations involved applying a map op- experience with parallel and distributed systems to eas- eration to each logical "record"in our input in order to ily utilize the resources of a large distributed system. compute a set of intermediate key/value pairs,and then Our implementation of MapReduce runs on a large applying a reduce operation to all the values that shared cluster of commodity machines and is highly scalable: the same key,in order to combine the derived data ap a typical MapReduce computation processes many ter- propriately.Our use of a functional model with user- abytes of data on thousands of machines.Programmers specified map and reduce operations allows us to paral- find the system easy to use:hundreds of MapReduce pro- lelize large computations easily and to use re-execution grams have been implemented and upwards of one thou- as the primary mechanism for fault tolerance. sand MapReduce jobs are executed on Google's clusters The major contributions of this work are a simple and every day. powerful interface that enables automatic parallelization and distribution of large-scale computations,combined with an implementation of this interface that achieves 1 Introduction high performance on large clusters of commodity PCs. Section 2 describes the basic programming model and Over the past five years,the authors and many others at gives several examples.Section 3 describes an imple- Google have implemented hundreds of special-purpose mentation of the MapReduce interface tailored towards computations that process large amounts of raw data, our cluster-based computing environment.Section 4 de- such as crawled documents,web request logs,etc.,to scribes several refinements of the programming model compute various kinds of derived data,such as inverted that we have found useful.Section 5 has performance indices,various representations of the graph structure measurements of our implementation for a variety of of web documents,summaries of the number of pages tasks.Section 6 explores the use of MapReduce within crawled per host,the set of most frequent queries in a Google including our experiences in using it as the basis To appear in OSDI 2004 1

MapReduce: Simplified Data Processing on Large Clusters Jeffrey Dean and Sanjay Ghemawat jeff@google.com, sanjay@google.com Google, Inc. Abstract MapReduce is a programming model and an associ￾ated implementation for processing and generating large data sets. Users specify a map function that processes a key/value pair to generate a set of intermediate key/value pairs, and a reduce function that merges all intermediate values associated with the same intermediate key. Many real world tasks are expressible in this model, as shown in the paper. Programs written in this functional style are automati￾cally parallelized and executed on a large cluster of com￾modity machines. The run-time system takes care of the details of partitioning the input data, scheduling the pro￾gram’s execution across a set of machines, handling ma￾chine failures, and managing the required inter-machine communication. This allows programmers without any experience with parallel and distributed systems to eas￾ily utilize the resources of a large distributed system. Our implementation of MapReduce runs on a large cluster of commodity machines and is highly scalable: a typical MapReduce computation processes many ter￾abytes of data on thousands of machines. Programmers find the system easy to use: hundreds of MapReduce pro￾grams have been implemented and upwards of one thou￾sand MapReduce jobs are executed on Google’s clusters every day. 1 Introduction Over the past five years, the authors and many others at Google have implemented hundreds of special-purpose computations that process large amounts of raw data, such as crawled documents, web request logs, etc., to compute various kinds of derived data, such as inverted indices, various representations of the graph structure of web documents, summaries of the number of pages crawled per host, the set of most frequent queries in a given day, etc. Most such computations are conceptu￾ally straightforward. However, the input data is usually large and the computations have to be distributed across hundreds or thousands of machines in order to finish in a reasonable amount of time. The issues of how to par￾allelize the computation, distribute the data, and handle failures conspire to obscure the original simple compu￾tation with large amounts of complex code to deal with these issues. As a reaction to this complexity, we designed a new abstraction that allows us to express the simple computa￾tions we were trying to perform but hides the messy de￾tails of parallelization, fault-tolerance, data distribution and load balancing in a library. Our abstraction is in￾spired by the map and reduce primitives present in Lisp and many other functional languages. We realized that most of our computations involved applying a map op￾eration to each logical “record” in our input in order to compute a set of intermediate key/value pairs, and then applying a reduce operation to all the values that shared the same key, in order to combine the derived data ap￾propriately. Our use of a functional model with user￾specified map and reduce operations allows us to paral￾lelize large computations easily and to use re-execution as the primary mechanism for fault tolerance. The major contributions of this work are a simple and powerful interface that enables automatic parallelization and distribution of large-scale computations, combined with an implementation of this interface that achieves high performance on large clusters of commodity PCs. Section 2 describes the basic programming model and gives several examples. Section 3 describes an imple￾mentation of the MapReduce interface tailored towards our cluster-based computing environment. Section 4 de￾scribes several refinements of the programming model that we have found useful. Section 5 has performance measurements of our implementation for a variety of tasks. Section 6 explores the use of MapReduce within Google including our experiences in using it as the basis To appear in OSDI 2004 1

for a rewrite of our production indexing system.Sec-2.2 Types tion 7 discusses related and future work. Even though the previous pseudo-code is written in terms 2 Programming Model of string inputs and outputs,conceptually the map and reduce functions supplied by the user have associated types: The computation takes a set of input key/value pairs,and produces a set of output key/value pairs.The user of map (k1,v1) →1ist(k2,v2) the MapReduce library expresses the computation as two reduce (k2,list (v2))-list(v2) functions:Map and Reduce. I.e.,the input keys and values are drawn from a different Map,written by the user,takes an input pair and pro- domain than the output keys and values.Furthermore. duces a set of intermediate key/value pairs.The MapRe- the intermediate keys and values are from the same do- duce library groups together all intermediate values asso- main as the output keys and values. ciated with the same intermediate key I and passes them Our C++implementation passes strings to and from to the Reduce function. the user-defined functions and leaves it to the user code The Reduce function,also written by the user,accepts to convert between strings and appropriate types. an intermediate key I and a set of values for that key.It merges together these values to form a possibly smaller set of values.Typically just zero or one output value is 2.3 More Examples produced per Reduce invocation.The intermediate val- ues are supplied to the user's reduce function via an iter- Here are a few simple examples of interesting programs ator.This allows us to handle lists of values that are too that can be easily expressed as MapReduce computa- large to fit in memory. tions. 2.1 Example Distributed Grep:The map function emits a line if it Consider the problem of counting the number of oc- matches a supplied pattern.The reduce function is an identity function that just copies the supplied intermedi- currences of each word in a large collection of docu- ate data to the output. ments.The user would write code similar to the follow- ing pseudo-code: Count of URL Access Frequency:The map func- map(String key,String value): /key:document name tion processes logs of web page requests and outputs /value:document contents (URL,1).The reduce function adds together all values for each word w in value: for the same URL and emits a (URL,total count) EmitIntermediate(w,"1"); pair. reduce(String key,Iterator values): /key:a word Reverse Web-Link Graph:The map function outputs /values:a list of counts (target,source)pairs for each link to a target int result =0; URL found in a page named source.The reduce for each v in values: function concatenates the list of all source URLs as- result +ParseInt(v); sociated with a given target URL and emits the pair: Emit (AsString(result)); (target,list(source)) The map function emits each word plus an associated count of occurrences (just'1'in this simple example) Term-Vector per Host:A term vector summarizes the The reduce function sums together all counts emitted most important words that occur in a document or a set for a particular word. of documents as a list of (word,frequency)pairs.The In addition,the user writes code to fill in a mapreduce map function emits a (hostname,term vector) specification object with the names of the input and out- pair for each input document (where the hostname is put files,and optional tuning parameters.The user then extracted from the URL of the document).The re- invokes the MapReduce function,passing it the specifi- duce function is passed all per-document term vectors cation object.The user's code is linked together with the for a given host.It adds these term vectors together. MapReduce library (implemented in C++).Appendix A throwing away infrequent terms,and then emits a final contains the full program text for this example. (hostname,term vector)pair. To appear in OSDI 2004 2

for a rewrite of our production indexing system. Sec￾tion 7 discusses related and future work. 2 Programming Model The computation takes a set of input key/value pairs, and produces a set of output key/value pairs. The user of the MapReduce library expresses the computation as two functions: Map and Reduce. Map, written by the user, takes an input pair and pro￾duces a set of intermediate key/value pairs. The MapRe￾duce library groups together all intermediate values asso￾ciated with the same intermediate key I and passes them to the Reduce function. The Reduce function, also written by the user, accepts an intermediate key I and a set of values for that key. It merges together these values to form a possibly smaller set of values. Typically just zero or one output value is produced per Reduce invocation. The intermediate val￾ues are supplied to the user’s reduce function via an iter￾ator. This allows us to handle lists of values that are too large to fit in memory. 2.1 Example Consider the problem of counting the number of oc￾currences of each word in a large collection of docu￾ments. The user would write code similar to the follow￾ing pseudo-code: map(String key, String value): // key: document name // value: document contents for each word w in value: EmitIntermediate(w, "1"); reduce(String key, Iterator values): // key: a word // values: a list of counts int result = 0; for each v in values: result += ParseInt(v); Emit(AsString(result)); The map function emits each word plus an associated count of occurrences (just ‘1’ in this simple example). The reduce function sums together all counts emitted for a particular word. In addition, the user writes code to fill in a mapreduce specification object with the names of the input and out￾put files, and optional tuning parameters. The user then invokes the MapReduce function, passing it the specifi- cation object. The user’s code is linked together with the MapReduce library (implemented in C++). Appendix A contains the full program text for this example. 2.2 Types Even though the previous pseudo-code is written in terms of string inputs and outputs, conceptually the map and reduce functions supplied by the user have associated types: map (k1,v1) → list(k2,v2) reduce (k2,list(v2)) → list(v2) I.e., the input keys and values are drawn from a different domain than the output keys and values. Furthermore, the intermediate keys and values are from the same do￾main as the output keys and values. Our C++ implementation passes strings to and from the user-defined functions and leaves it to the user code to convert between strings and appropriate types. 2.3 More Examples Here are a few simple examples of interesting programs that can be easily expressed as MapReduce computa￾tions. Distributed Grep: The map function emits a line if it matches a supplied pattern. The reduce function is an identity function that just copies the supplied intermedi￾ate data to the output. Count of URL Access Frequency: The map func￾tion processes logs of web page requests and outputs hURL, 1i. The reduce function adds together all values for the same URL and emits a hURL, total counti pair. Reverse Web-Link Graph: The map function outputs htarget, sourcei pairs for each link to a target URL found in a page named source. The reduce function concatenates the list of all source URLs as￾sociated with a given target URL and emits the pair: htarget, list(source)i Term-Vector per Host: A term vector summarizes the most important words that occur in a document or a set of documents as a list of hword, frequencyi pairs. The map function emits a hhostname, term vectori pair for each input document (where the hostname is extracted from the URL of the document). The re￾duce function is passed all per-document term vectors for a given host. It adds these term vectors together, throwing away infrequent terms, and then emits a final hhostname, term vectori pair. To appear in OSDI 2004 2

User Program (1)fork (1)fork (1)fork Master 2) (2) assign assign reduce. 、 map worker split 0 (6)write output split 1 worker (5)remote read file 0 split 2 (3)read (4)local write worker split 3 worker output file 1 split 4 worker Input Map Intermediate files Reduce Output files phase (on local disks) phase files Figure 1:Execution overview Inverted Index:The map function parses each docu- large clusters of commodity PCs connected together with ment,and emits a sequence of (word,document ID) switched Ethernet [4].In our environment: pairs.The reduce function accepts all pairs for a given (1)Machines are typically dual-processor x86 processors word,sorts the corresponding document IDs and emits a running Linux,with 2-4 GB of memory per machine. (word,list(document ID))pair.The set of all output pairs forms a simple inverted index.It is easy to augment (2)Commodity networking hardware is used-typically this computation to keep track of word positions. either 100 megabits/second or 1 gigabit/second at the machine level,but averaging considerably less in over- all bisection bandwidth. Distributed Sort:The map function extracts the key (3)A cluster consists of hundreds or thousands of ma- from each record,and emits a(key,record)pair.The chines.and therefore machine failures are common. reduce function emits all pairs unchanged.This compu- tation depends on the partitioning facilities described in (4)Storage is provided by inexpensive IDE disks at- Section 4.1 and the ordering properties described in Sec- tached directly to individual machines.A distributed file tion 4.2 system [8]developed in-house is used to manage the data stored on these disks.The file system uses replication to provide availability and reliability on top of unreliable 3 Implementation hardware. (5)Users submit jobs to a scheduling system.Each job Many different implementations of the MapReduce in- terface are possible.The right choice depends on the consists of a set of tasks,and is mapped by the scheduler to a set of available machines within a cluster. environment.For example,one implementation may be suitable for a small shared-memory machine,another for a large NUMA multi-processor,and yet another for an 3.1 Execution Overview even larger collection of networked machines This section describes an implementation targeted The Map invocations are distributed across multiple to the computing environment in wide use at Google: machines by automatically partitioning the input data To appear in OSDI 2004 3

User Program Master (1) fork worker (1) fork worker (1) fork (2) assign map (2) assign reduce split 0 split 1 split 2 split 3 split 4 output file 0 (6) write worker (3) read worker (4) local write Map phase Intermediate files (on local disks) worker output file 1 Input files (5) remote read Reduce phase Output files Figure 1: Execution overview Inverted Index: The map function parses each docu￾ment, and emits a sequence of hword, document IDi pairs. The reduce function accepts all pairs for a given word, sorts the corresponding document IDs and emits a hword, list(document ID)i pair. The set of all output pairs forms a simple inverted index. It is easy to augment this computation to keep track of word positions. Distributed Sort: The map function extracts the key from each record, and emits a hkey, recordi pair. The reduce function emits all pairs unchanged. This compu￾tation depends on the partitioning facilities described in Section 4.1 and the ordering properties described in Sec￾tion 4.2. 3 Implementation Many different implementations of the MapReduce in￾terface are possible. The right choice depends on the environment. For example, one implementation may be suitable for a small shared-memory machine, another for a large NUMA multi-processor, and yet another for an even larger collection of networked machines. This section describes an implementation targeted to the computing environment in wide use at Google: large clusters of commodity PCs connected together with switched Ethernet [4]. In our environment: (1) Machines are typically dual-processor x86 processors running Linux, with 2-4 GB of memory per machine. (2) Commodity networking hardware is used – typically either 100 megabits/second or 1 gigabit/second at the machine level, but averaging considerably less in over￾all bisection bandwidth. (3) A cluster consists of hundreds or thousands of ma￾chines, and therefore machine failures are common. (4) Storage is provided by inexpensive IDE disks at￾tached directly to individual machines. A distributed file system [8] developed in-house is used to manage the data stored on these disks. The file system uses replication to provide availability and reliability on top of unreliable hardware. (5) Users submit jobs to a scheduling system. Each job consists of a set of tasks, and is mapped by the scheduler to a set of available machines within a cluster. 3.1 Execution Overview The Map invocations are distributed across multiple machines by automatically partitioning the input data To appear in OSDI 2004 3

into a set of M splits.The input splits can be pro- 7.When all map tasks and reduce tasks have been cessed in parallel by different machines.Reduce invoca- completed,the master wakes up the user program. tions are distributed by partitioning the intermediate key At this point,the MapReduce call in the user pro- space into R pieces using a partitioning function (e.g., gram returns back to the user code. hash(key)mod R).The number of partitions(R)and the partitioning function are specified by the user. After successful completion,the output of the mapre- Figure 1 shows the overall flow of a MapReduce op- duce execution is available in the R output files (one per eration in our implementation.When the user program reduce task,with file names as specified by the user). calls the MapReduce function,the following sequence Typically,users do not need to combine these R output of actions occurs(the numbered labels in Figure 1 corre- files into one file-they often pass these files as input to spond to the numbers in the list below): another MapReduce call,or use them from another dis- tributed application that is able to deal with input that is 1.The MapReduce library in the user program first partitioned into multiple files. splits the input files into M pieces of typically 16 megabytes to 64 megabytes (MB)per piece (con- 3.2 Master Data Structures trollable by the user via an optional parameter).It then starts up many copies of the program on a clus- The master keeps several data structures.For each map ter of machines. task and reduce task,it stores the state (idle,in-progress, 2.One of the copies of the program is special-the or completed),and the identity of the worker machine (for non-idle tasks). master.The rest are workers that are assigned work by the master.There are M map tasks and R reduce The master is the conduit through which the location tasks to assign.The master picks idle workers and of intermediate file regions is propagated from map tasks assigns each one a map task or a reduce task. to reduce tasks.Therefore,for each completed map task. the master stores the locations and sizes of the R inter- 3.A worker who is assigned a map task reads the mediate file regions produced by the map task.Updates contents of the corresponding input split.It parses to this location and size information are received as map key/value pairs out of the input data and passes each tasks are completed.The information is pushed incre- pair to the user-defined Map function.The interme- mentally to workers that have in-progress reduce tasks. diate key/value pairs produced by the Map function are buffered in memory. 3.3 Fault Tolerance 4.Periodically,the buffered pairs are written to local disk,partitioned into R regions by the partitioning Since the MapReduce library is designed to help process function.The locations of these buffered pairs on very large amounts of data using hundreds or thousands the local disk are passed back to the master,who of machines,the library must tolerate machine failures is responsible for forwarding these locations to the gracefully. reduce workers. Worker Failure 5.When a reduce worker is notified by the master about these locations,it uses remote procedure calls The master pings every worker periodically.If no re- to read the buffered data from the local disks of the sponse is received from a worker in a certain amount of map workers.When a reduce worker has read all in- time,the master marks the worker as failed.Any map termediate data,it sorts it by the intermediate keys tasks completed by the worker are reset back to their ini- so that all occurrences of the same key are grouped tial idle state,and therefore become eligible for schedul- together.The sorting is needed because typically ing on other workers.Similarly,any map task or reduce many different keys map to the same reduce task.If task in progress on a failed worker is also reset to idle the amount of intermediate data is too large to fit in and becomes eligible for rescheduling. memory,an external sort is used. Completed map tasks are re-executed on a failure be- 6.The reduce worker iterates over the sorted interme- cause their output is stored on the local disk(s)of the diate data and for each unique intermediate key en- failed machine and is therefore inaccessible.Completed countered,it passes the key and the corresponding reduce tasks do not need to be re-executed since their set of intermediate values to the user's Reduce func- output is stored in a global file system tion.The output of the Reduce function is appended When a map task is executed first by worker A and to a final output file for this reduce partition. then later executed by worker B (because A failed).all To appear in OSDI 2004 4

into a set of M splits. The input splits can be pro￾cessed in parallel by different machines. Reduce invoca￾tions are distributed by partitioning the intermediate key space into R pieces using a partitioning function (e.g., hash(key) mod R). The number of partitions (R) and the partitioning function are specified by the user. Figure 1 shows the overall flow of a MapReduce op￾eration in our implementation. When the user program calls the MapReduce function, the following sequence of actions occurs (the numbered labels in Figure 1 corre￾spond to the numbers in the list below): 1. The MapReduce library in the user program first splits the input files into M pieces of typically 16 megabytes to 64 megabytes (MB) per piece (con￾trollable by the user via an optional parameter). It then starts up many copies of the program on a clus￾ter of machines. 2. One of the copies of the program is special – the master. The rest are workers that are assigned work by the master. There are M map tasks and R reduce tasks to assign. The master picks idle workers and assigns each one a map task or a reduce task. 3. A worker who is assigned a map task reads the contents of the corresponding input split. It parses key/value pairs out of the input data and passes each pair to the user-defined Map function. The interme￾diate key/value pairs produced by the Map function are buffered in memory. 4. Periodically, the buffered pairs are written to local disk, partitioned into R regions by the partitioning function. The locations of these buffered pairs on the local disk are passed back to the master, who is responsible for forwarding these locations to the reduce workers. 5. When a reduce worker is notified by the master about these locations, it uses remote procedure calls to read the buffered data from the local disks of the map workers. When a reduce worker has read all in￾termediate data, it sorts it by the intermediate keys so that all occurrences of the same key are grouped together. The sorting is needed because typically many different keys map to the same reduce task. If the amount of intermediate data is too large to fit in memory, an external sort is used. 6. The reduce worker iterates over the sorted interme￾diate data and for each unique intermediate key en￾countered, it passes the key and the corresponding set of intermediate values to the user’s Reduce func￾tion. The output of the Reduce function is appended to a final output file for this reduce partition. 7. When all map tasks and reduce tasks have been completed, the master wakes up the user program. At this point, the MapReduce call in the user pro￾gram returns back to the user code. After successful completion, the output of the mapre￾duce execution is available in the R output files (one per reduce task, with file names as specified by the user). Typically, users do not need to combine these R output files into one file – they often pass these files as input to another MapReduce call, or use them from another dis￾tributed application that is able to deal with input that is partitioned into multiple files. 3.2 Master Data Structures The master keeps several data structures. For each map task and reduce task, it stores the state (idle, in-progress, or completed), and the identity of the worker machine (for non-idle tasks). The master is the conduit through which the location of intermediate file regions is propagated from map tasks to reduce tasks. Therefore, for each completed map task, the master stores the locations and sizes of the R inter￾mediate file regions produced by the map task. Updates to this location and size information are received as map tasks are completed. The information is pushed incre￾mentally to workers that have in-progress reduce tasks. 3.3 Fault Tolerance Since the MapReduce library is designed to help process very large amounts of data using hundreds or thousands of machines, the library must tolerate machine failures gracefully. Worker Failure The master pings every worker periodically. If no re￾sponse is received from a worker in a certain amount of time, the master marks the worker as failed. Any map tasks completed by the worker are reset back to their ini￾tial idle state, and therefore become eligible for schedul￾ing on other workers. Similarly, any map task or reduce task in progress on a failed worker is also reset to idle and becomes eligible for rescheduling. Completed map tasks are re-executed on a failure be￾cause their output is stored on the local disk(s) of the failed machine and is therefore inaccessible. Completed reduce tasks do not need to be re-executed since their output is stored in a global file system. When a map task is executed first by worker A and then later executed by worker B (because A failed), all To appear in OSDI 2004 4

workers executing reduce tasks are notified of the re- easy for programmers to reason about their program's be- execution.Any reduce task that has not already read the havior.When the map and/or reduce operators are non- data from worker A will read the data from worker B. deterministic,we provide weaker but still reasonable se- MapReduce is resilient to large-scale worker failures. mantics.In the presence of non-deterministic operators. For example,during one MapReduce operation,network the output of a particular reduce task R is equivalent to maintenance on a running cluster was causing groups of the output for Ri produced by a sequential execution of 80 machines at a time to become unreachable for sev- the non-deterministic program.However,the output for eral minutes.The MapReduce master simply re-executed a different reduce task R2 may correspond to the output the work done by the unreachable worker machines,and for R2 produced by a different sequential execution of continued to make forward progress,eventually complet- the non-deterministic program. ing the MapReduce operation. Consider map task M and reduce tasks Ri and R2. Let e(R;)be the execution of R,that committed (there is exactly one such execution).The weaker semantics Master Failure arise because e(R)may have read the output produced by one execution of M and e(R2)may have read the It is easy to make the master write periodic checkpoints output produced by a different execution of M. of the master data structures described above.If the mas- ter task dies,a new copy can be started from the last 3.4 Locality checkpointed state.However,given that there is only a single master,its failure is unlikely;therefore our cur- Network bandwidth is a relatively scarce resource in our rent implementation aborts the MapReduce computation computing environment.We conserve network band- if the master fails.Clients can check for this condition width by taking advantage of the fact that the input data and retry the MapReduce operation if they desire. (managed by GFS [8])is stored on the local disks of the machines that make up our cluster.GFS divides each file into 64 MB blocks,and stores several copies of each Semantics in the Presence of Failures block (typically 3 copies)on different machines.The MapReduce master takes the location information of the When the user-supplied map and reduce operators are de- input files into account and attempts to schedule a map terministic functions of their input values,our distributed task on a machine that contains a replica of the corre- implementation produces the same output as would have sponding input data.Failing that,it attempts to schedule been produced by a non-faulting sequential execution of a map task near a replica of that task's input data(e.g.,on the entire program. a worker machine that is on the same network switch as We rely on atomic commits of map and reduce task the machine containing the data).When running large outputs to achieve this property.Each in-progress task MapReduce operations on a significant fraction of the writes its output to private temporary files.A reduce task workers in a cluster,most input data is read locally and produces one such file,and a map task produces R such consumes no network bandwidth. files (one per reduce task).When a map task completes, the worker sends a message to the master and includes 3.5 Task Granularity the names of the R temporary files in the message.If the master receives a completion message for an already We subdivide the map phase into M pieces and the re- completed map task.it ignores the message.Otherwise. duce phase into R pieces,as described above.Ideally,M it records the names of R files in a master data structure. and R should be much larger than the number of worker machines.Having each worker perform many different When a reduce task completes,the reduce worker tasks improves dynamic load balancing,and also speeds atomically renames its temporary output file to the final up recovery when a worker fails:the many map tasks output file.If the same reduce task is executed on multi- it has completed can be spread out across all the other ple machines,multiple rename calls will be executed for worker machines. the same final output file.We rely on the atomic rename There are practical bounds on how large M and R can operation provided by the underlying file system to guar- be in our implementation.since the master must make antee that the final file system state contains just the data O(M+R)scheduling decisions and keeps O(M R) produced by one execution of the reduce task. state in memory as described above.(The constant fac- The vast majority of our map and reduce operators are tors for memory usage are small however:the O(M R) deterministic,and the fact that our semantics are equiv- piece of the state consists of approximately one byte of alent to a sequential execution in this case makes it very data per map task/reduce task pair.) To appear in OSDI 2004

workers executing reduce tasks are notified of the re￾execution. Any reduce task that has not already read the data from worker A will read the data from worker B. MapReduce is resilient to large-scale worker failures. For example, during one MapReduce operation, network maintenance on a running cluster was causing groups of 80 machines at a time to become unreachable for sev￾eral minutes. The MapReduce mastersimply re-executed the work done by the unreachable worker machines, and continued to make forward progress, eventually complet￾ing the MapReduce operation. Master Failure It is easy to make the master write periodic checkpoints of the master data structures described above. If the mas￾ter task dies, a new copy can be started from the last checkpointed state. However, given that there is only a single master, its failure is unlikely; therefore our cur￾rent implementation aborts the MapReduce computation if the master fails. Clients can check for this condition and retry the MapReduce operation if they desire. Semantics in the Presence of Failures When the user-supplied map and reduce operators are de￾terministic functions of their input values, our distributed implementation produces the same output as would have been produced by a non-faulting sequential execution of the entire program. We rely on atomic commits of map and reduce task outputs to achieve this property. Each in-progress task writes its output to private temporary files. A reduce task produces one such file, and a map task produces R such files (one per reduce task). When a map task completes, the worker sends a message to the master and includes the names of the R temporary files in the message. If the master receives a completion message for an already completed map task, it ignores the message. Otherwise, it records the names of R files in a master data structure. When a reduce task completes, the reduce worker atomically renames its temporary output file to the final output file. If the same reduce task is executed on multi￾ple machines, multiple rename calls will be executed for the same final output file. We rely on the atomic rename operation provided by the underlying file system to guar￾antee that the final file system state contains just the data produced by one execution of the reduce task. The vast majority of our map and reduce operators are deterministic, and the fact that our semantics are equiv￾alent to a sequential execution in this case makes it very easy for programmersto reason about their program’s be￾havior. When the map and/or reduce operators are non￾deterministic, we provide weaker but still reasonable se￾mantics. In the presence of non-deterministic operators, the output of a particular reduce task R1 is equivalent to the output for R1 produced by a sequential execution of the non-deterministic program. However, the output for a different reduce task R2 may correspond to the output for R2 produced by a different sequential execution of the non-deterministic program. Consider map task M and reduce tasks R1 and R2. Let e(Ri) be the execution of Ri that committed (there is exactly one such execution). The weaker semantics arise because e(R1) may have read the output produced by one execution of M and e(R2) may have read the output produced by a different execution of M. 3.4 Locality Network bandwidth is a relatively scarce resource in our computing environment. We conserve network band￾width by taking advantage of the fact that the input data (managed by GFS [8]) is stored on the local disks of the machines that make up our cluster. GFS divides each file into 64 MB blocks, and stores several copies of each block (typically 3 copies) on different machines. The MapReduce master takes the location information of the input files into account and attempts to schedule a map task on a machine that contains a replica of the corre￾sponding input data. Failing that, it attempts to schedule a map task near a replica of that task’s input data (e.g., on a worker machine that is on the same network switch as the machine containing the data). When running large MapReduce operations on a significant fraction of the workers in a cluster, most input data is read locally and consumes no network bandwidth. 3.5 Task Granularity We subdivide the map phase into M pieces and the re￾duce phase into R pieces, as described above. Ideally, M and R should be much larger than the number of worker machines. Having each worker perform many different tasks improves dynamic load balancing, and also speeds up recovery when a worker fails: the many map tasks it has completed can be spread out across all the other worker machines. There are practical bounds on how large M and R can be in our implementation, since the master must make O(M + R) scheduling decisions and keeps O(M ∗ R) state in memory as described above. (The constant fac￾tors for memory usage are small however: the O(M ∗R) piece of the state consists of approximately one byte of data per map task/reduce task pair.) To appear in OSDI 2004 5

Furthermore,R is often constrained by users because the intermediate key.A default partitioning function is the output of each reduce task ends up in a separate out- provided that uses hashing (e.g."hash(key)mod R"). put file.In practice,we tend to choose M so that each This tends to result in fairly well-balanced partitions.In individual task is roughly 16 MB to 64 MB of input data some cases,however,it is useful to partition data by (so that the locality optimization described above is most some other function of the key.For example,sometimes effective),and we make R a small multiple of the num- the output keys are URLs,and we want all entries for a ber of worker machines we expect to use.We often per- single host to end up in the same output file.To support form MapReduce computations with M=200,000 and situations like this,the user of the MapReduce library R=5.000,using 2.000 worker machines. can provide a special partitioning function.For example. using "hash(Hostname(urlkey))mod R"as the par- titioning function causes all URLs from the same host to 3.6 Backup Tasks end up in the same output file One of the common causes that lengthens the total time taken for a MapReduce operation is a"straggler":a ma- 4.2 Ordering Guarantees chine that takes an unusually long time to complete one of the last few map or reduce tasks in the computation. We guarantee that within a given partition,the interme- Stragglers can arise for a whole host of reasons.For ex- diate key/value pairs are processed in increasing key or- ample,a machine with a bad disk may experience fre- der.This ordering guarantee makes it easy to generate quent correctable errors that slow its read performance a sorted output file per partition,which is useful when from 30 MB/s to 1 MB/s.The cluster scheduling sys- the output file format needs to support efficient random tem may have scheduled other tasks on the machine, access lookups by key,or users of the output find it con- causing it to execute the MapReduce code more slowly venient to have the data sorted. due to competition for CPU,memory,local disk,or net- work bandwidth.A recent problem we experienced was 4.3 Combiner Function a bug in machine initialization code that caused proces- sor caches to be disabled:computations on affected ma- In some cases,there is significant repetition in the inter- chines slowed down by over a factor of one hundred. mediate keys produced by each map task,and the user- We have a general mechanism to alleviate the prob- specified Reduce function is commutative and associa- lem of stragglers.When a MapReduce operation is close tive.A good example of this is the word counting exam- to completion,the master schedules backup executions ple in Section 2.1.Since word frequencies tend to follow of the remaining in-progress tasks.The task is marked a Zipf distribution,each map task will produce hundreds as completed whenever either the primary or the backup or thousands of records of the form .All of execution completes.We have tuned this mechanism so these counts will be sent over the network to a single re- that it typically increases the computational resources duce task and then added together by the Reduce function used by the operation by no more than a few percent. to produce one number.We allow the user to specify an We have found that this significantly reduces the time optional Combiner function that does partial merging of to complete large MapReduce operations.As an exam- this data before it is sent over the network. ple,the sort program described in Section 5.3 takes 44% The Combiner function is executed on each machine longer to complete when the backup task mechanism is that performs a map task.Typically the same code is used disabled. to implement both the combiner and the reduce func- tions.The only difference between a reduce function and a combiner function is how the MapReduce library han- 4 Refinements dles the output of the function.The output of a reduce function is written to the final output file.The output of Although the basic functionality provided by simply a combiner function is written to an intermediate file that writing Map and Reduce functions is sufficient for most will be sent to a reduce task. needs.we have found a few extensions useful.These are Partial combining significantly speeds up certain described in this section. classes of MapReduce operations.Appendix A contains an example that uses a combiner. 4.1 Partitioning Function 4.4 Input and Output Types The users of MapReduce specify the number of reduce tasks/output files that they desire(R).Data gets parti- The MapReduce library provides support for reading in- tioned across these tasks using a partitioning function on put data in several different formats.For example,"text" To appear in OSDI 2004 6

Furthermore, R is often constrained by users because the output of each reduce task ends up in a separate out￾put file. In practice, we tend to choose M so that each individual task is roughly 16 MB to 64 MB of input data (so that the locality optimization described above is most effective), and we make R a small multiple of the num￾ber of worker machines we expect to use. We often per￾form MapReduce computations with M = 200, 000 and R = 5, 000, using 2,000 worker machines. 3.6 Backup Tasks One of the common causes that lengthens the total time taken for a MapReduce operation is a “straggler”: a ma￾chine that takes an unusually long time to complete one of the last few map or reduce tasks in the computation. Stragglers can arise for a whole host of reasons. For ex￾ample, a machine with a bad disk may experience fre￾quent correctable errors that slow its read performance from 30 MB/s to 1 MB/s. The cluster scheduling sys￾tem may have scheduled other tasks on the machine, causing it to execute the MapReduce code more slowly due to competition for CPU, memory, local disk, or net￾work bandwidth. A recent problem we experienced was a bug in machine initialization code that caused proces￾sor caches to be disabled: computations on affected ma￾chines slowed down by over a factor of one hundred. We have a general mechanism to alleviate the prob￾lem of stragglers. When a MapReduce operation is close to completion, the master schedules backup executions of the remaining in-progress tasks. The task is marked as completed whenever either the primary or the backup execution completes. We have tuned this mechanism so that it typically increases the computational resources used by the operation by no more than a few percent. We have found that this significantly reduces the time to complete large MapReduce operations. As an exam￾ple, the sort program described in Section 5.3 takes 44% longer to complete when the backup task mechanism is disabled. 4 Refinements Although the basic functionality provided by simply writing Map and Reduce functions is sufficient for most needs, we have found a few extensions useful. These are described in this section. 4.1 Partitioning Function The users of MapReduce specify the number of reduce tasks/output files that they desire (R). Data gets parti￾tioned across these tasks using a partitioning function on the intermediate key. A default partitioning function is provided that uses hashing (e.g. “hash(key) mod R”). This tends to result in fairly well-balanced partitions. In some cases, however, it is useful to partition data by some other function of the key. For example, sometimes the output keys are URLs, and we want all entries for a single host to end up in the same output file. To support situations like this, the user of the MapReduce library can provide a special partitioning function. For example, using “hash(Hostname(urlkey)) mod R” as the par￾titioning function causes all URLs from the same host to end up in the same output file. 4.2 Ordering Guarantees We guarantee that within a given partition, the interme￾diate key/value pairs are processed in increasing key or￾der. This ordering guarantee makes it easy to generate a sorted output file per partition, which is useful when the output file format needs to support efficient random access lookups by key, or users of the output find it con￾venient to have the data sorted. 4.3 Combiner Function In some cases, there is significant repetition in the inter￾mediate keys produced by each map task, and the user￾specified Reduce function is commutative and associa￾tive. A good example of this is the word counting exam￾ple in Section 2.1. Since word frequencies tend to follow a Zipf distribution, each map task will produce hundreds or thousands of records of the form . All of these counts will be sent over the network to a single re￾duce task and then added together by the Reduce function to produce one number. We allow the user to specify an optional Combiner function that does partial merging of this data before it is sent over the network. The Combiner function is executed on each machine that performs a map task. Typically the same code is used to implement both the combiner and the reduce func￾tions. The only difference between a reduce function and a combiner function is how the MapReduce library han￾dles the output of the function. The output of a reduce function is written to the final output file. The output of a combiner function is written to an intermediate file that will be sent to a reduce task. Partial combining significantly speeds up certain classes of MapReduce operations. Appendix A contains an example that uses a combiner. 4.4 Input and Output Types The MapReduce library provides support for reading in￾put data in several different formats. For example, “text” To appear in OSDI 2004 6

mode input treats each line as a key/value pair:the key the signal handler sends a "last gasp"UDP packet that is the offset in the file and the value is the contents of contains the sequence number to the MapReduce mas- the line.Another common supported format stores a ter.When the master has seen more than one failure on sequence of key/value pairs sorted by key.Each input a particular record,it indicates that the record should be type implementation knows how to split itself into mean- skipped when it issues the next re-execution of the corre- ingful ranges for processing as separate map tasks(e.g. sponding Map or Reduce task text mode's range splitting ensures that range splits oc- cur only at line boundaries).Users can add support for a new input type by providing an implementation of a sim- 4.7 Local Execution ple reader interface,though most users just use one of a small number of predefined input types. Debugging problems in Map or Reduce functions can be A reader does not necessarily need to provide data tricky,since the actual computation happens in a dis- tributed system,often on several thousand machines, read from a file.For example,it is easy to define a reader with work assignment decisions made dynamically by that reads records from a database,or from data struc- the master.To help facilitate debugging,profiling,and tures mapped in memory. small-scale testing,we have developed an alternative im- In a similar fashion,we support a set of output types plementation of the MapReduce library that sequentially for producing data in different formats and it is easy for executes all of the work for a MapReduce operation on user code to add support for new output types. the local machine.Controls are provided to the user so that the computation can be limited to particular map 4.5 Side-effects tasks.Users invoke their program with a special flag and can then easily use any debugging or testing tools they In some cases,users of MapReduce have found it con- find useful (e.g.gdb). venient to produce auxiliary files as additional outputs from their map and/or reduce operators.We rely on the application writer to make such side-effects atomic and 4.8 Status Information idempotent.Typically the application writes to a tempo- rary file and atomically renames this file once it has been The master runs an internal HTTP server and exports fully generated. a set of status pages for human consumption.The sta- tus pages show the progress of the computation,such as We do not provide support for atomic two-phase com- how many tasks have been completed,how many are in mits of multiple output files produced by a single task. progress,bytes of input,bytes of intermediate data,bytes Therefore,tasks that produce multiple output files with of output,processing rates,etc.The pages also contain cross-file consistency requirements should be determin- links to the standard error and standard output files gen- istic.This restriction has never been an issue in practice. erated by each task.The user can use this data to pre- dict how long the computation will take,and whether or 4.6 Skipping Bad Records not more resources should be added to the computation. These pages can also be used to figure out when the com- Sometimes there are bugs in user code that cause the Map putation is much slower than expected. or Reduce functions to crash deterministically on certain In addition,the top-level status page shows which records.Such bugs prevent a MapReduce operation from workers have failed,and which map and reduce tasks completing.The usual course of action is to fix the bug, they were processing when they failed.This informa- but sometimes this is not feasible;perhaps the bug is in tion is useful when attempting to diagnose bugs in the a third-party library for which source code is unavail- user code. able.Also,sometimes it is acceptable to ignore a few records,for example when doing statistical analysis on a large data set.We provide an optional mode of execu- 4.9 Counters tion where the MapReduce library detects which records cause deterministic crashes and skips these records in or- The MapReduce library provides a counter facility to der to make forward progress. count occurrences of various events.For example,user Each worker process installs a signal handler that code may want to count total number of words processed catches segmentation violations and bus errors.Before or the number of German documents indexed,etc. invoking a user Map or Reduce operation,the MapRe- To use this facility,user code creates a named counter duce library stores the sequence number of the argument object and then increments the counter appropriately in in a global variable.If the user code generates a signal, the Map and/or Reduce function.For example: To appear in OSDI 2004

mode input treats each line as a key/value pair: the key is the offset in the file and the value is the contents of the line. Another common supported format stores a sequence of key/value pairs sorted by key. Each input type implementation knows how to split itself into mean￾ingful ranges for processing as separate map tasks (e.g. text mode’s range splitting ensures that range splits oc￾cur only at line boundaries). Users can add support for a new input type by providing an implementation of a sim￾ple reader interface, though most users just use one of a small number of predefined input types. A reader does not necessarily need to provide data read from a file. For example, it is easy to define a reader that reads records from a database, or from data struc￾tures mapped in memory. In a similar fashion, we support a set of output types for producing data in different formats and it is easy for user code to add support for new output types. 4.5 Side-effects In some cases, users of MapReduce have found it con￾venient to produce auxiliary files as additional outputs from their map and/or reduce operators. We rely on the application writer to make such side-effects atomic and idempotent. Typically the application writes to a tempo￾rary file and atomically renames this file once it has been fully generated. We do not provide support for atomic two-phase com￾mits of multiple output files produced by a single task. Therefore, tasks that produce multiple output files with cross-file consistency requirements should be determin￾istic. This restriction has never been an issue in practice. 4.6 Skipping Bad Records Sometimesthere are bugsin user code that cause the Map or Reduce functions to crash deterministically on certain records. Such bugs prevent a MapReduce operation from completing. The usual course of action is to fix the bug, but sometimes this is not feasible; perhaps the bug is in a third-party library for which source code is unavail￾able. Also, sometimes it is acceptable to ignore a few records, for example when doing statistical analysis on a large data set. We provide an optional mode of execu￾tion where the MapReduce library detects which records cause deterministic crashes and skips these records in or￾der to make forward progress. Each worker process installs a signal handler that catches segmentation violations and bus errors. Before invoking a user Map or Reduce operation, the MapRe￾duce library stores the sequence number of the argument in a global variable. If the user code generates a signal, the signal handler sends a “last gasp” UDP packet that contains the sequence number to the MapReduce mas￾ter. When the master has seen more than one failure on a particular record, it indicates that the record should be skipped when it issues the next re-execution of the corre￾sponding Map or Reduce task. 4.7 Local Execution Debugging problems in Map or Reduce functions can be tricky, since the actual computation happens in a dis￾tributed system, often on several thousand machines, with work assignment decisions made dynamically by the master. To help facilitate debugging, profiling, and small-scale testing, we have developed an alternative im￾plementation of the MapReduce library that sequentially executes all of the work for a MapReduce operation on the local machine. Controls are provided to the user so that the computation can be limited to particular map tasks. Users invoke their program with a special flag and can then easily use any debugging or testing tools they find useful (e.g. gdb). 4.8 Status Information The master runs an internal HTTP server and exports a set of status pages for human consumption. The sta￾tus pages show the progress of the computation, such as how many tasks have been completed, how many are in progress, bytes of input, bytes of intermediate data, bytes of output, processing rates, etc. The pages also contain links to the standard error and standard output files gen￾erated by each task. The user can use this data to pre￾dict how long the computation will take, and whether or not more resources should be added to the computation. These pages can also be used to figure out when the com￾putation is much slower than expected. In addition, the top-level status page shows which workers have failed, and which map and reduce tasks they were processing when they failed. This informa￾tion is useful when attempting to diagnose bugs in the user code. 4.9 Counters The MapReduce library provides a counter facility to count occurrences of various events. For example, user code may want to count total number of words processed or the number of German documents indexed, etc. To use this facility, user code creates a named counter object and then increments the counter appropriately in the Map and/or Reduce function. For example: To appear in OSDI 2004 7

Counter*uppercase; 30000 uppercase GetCounter("uppercase"); 20000 map(String name,String contents): for each word w in contents: G二 10000 if (IsCapitalized(w)): uppercase->Increment () 0 EmitIntermediate(w,"1"); 20 40 60 80 100 Seconds The counter values from individual worker machines are periodically propagated to the master(piggybacked Figure 2:Data transfer rate over time on the ping response).The master aggregates the counter values from successful map and reduce tasks and returns them to the user code when the MapReduce operation disks,and a gigabit Ethernet link.The machines were is completed.The current counter values are also dis- arranged in a two-level tree-shaped switched network played on the master status page so that a human can with approximately 100-200 Gbps of aggregate band- watch the progress of the live computation.When aggre- width available at the root.All of the machines were gating counter values,the master eliminates the effects of in the same hosting facility and therefore the round-trip duplicate executions of the same map or reduce task to time between any pair of machines was less than a mil- avoid double counting.(Duplicate executions can arise lisecond. from our use of backup tasks and from re-execution of Out of the 4GB of memory,approximately 1-1.5GB tasks due to failures.) was reserved by other tasks running on the cluster.The Some counter values are automatically maintained programs were executed on a weekend afternoon.when by the MapReduce library,such as the number of in- the CPUs,disks,and network were mostly idle. put key/value pairs processed and the number of output key/value pairs produced. Users have found the counter facility useful for san- 5.2 Grep ity checking the behavior of MapReduce operations.For example,in some MapReduce operations,the user code The grep program scans through 1010 100-byte records may want to ensure that the number of output pairs searching for a relatively rare three-character pattern(the produced exactly equals the number of input pairs pro- pattern occurs in 92,337 records).The input is split into cessed,or that the fraction of German documents pro- approximately 64MB pieces(M=15000),and the en- cessed is within some tolerable fraction of the total num- tire output is placed in one file (R=1). ber of documents processed. Figure 2 shows the progress of the computation over time.The Y-axis shows the rate at which the input data is scanned.The rate gradually picks up as more machines 5 Performance are assigned to this MapReduce computation,and peaks In this section we measure the performance of MapRe- at over 30 GB/s when 1764 workers have been assigned. As the map tasks finish,the rate starts dropping and hits duce on two computations running on a large cluster of zero about 80 seconds into the computation.The entire machines.One computation searches through approxi- mately one terabyte of data looking for a particular pat- computation takes approximately 150 seconds from start to finish.This includes about a minute of startup over- tern.The other computation sorts approximately one ter- head.The overhead is due to the propagation of the pro- abyte of data. gram to all worker machines,and delays interacting with These two programs are representative of a large sub- GFS to open the set of 1000 input files and to get the set of the real programs written by users of MapReduce- information needed for the locality optimization. one class of programs shuffles data from one representa- tion to another,and another class extracts a small amount of interesting data from a large data set. 5.3 Sort 5.1 Cluster Configuration The sort program sorts 1010 100-byte records(approxi- mately 1 terabyte of data).This program is modeled after All of the programs were executed on a cluster that the TeraSort benchmark [10]. consisted of approximately 1800 machines.Each ma- The sorting program consists of less than 50 lines of chine had two 2GHz Intel Xeon processors with Hyper- user code.A three-line Map function extracts a 10-byte Threading enabled,4GB of memory,two 160GB IDE sorting key from a text line and emits the key and the To appear in OSDI 2004

Counter* uppercase; uppercase = GetCounter("uppercase"); map(String name, String contents): for each word w in contents: if (IsCapitalized(w)): uppercase->Increment(); EmitIntermediate(w, "1"); The counter values from individual worker machines are periodically propagated to the master (piggybacked on the ping response). The master aggregatesthe counter values from successful map and reduce tasks and returns them to the user code when the MapReduce operation is completed. The current counter values are also dis￾played on the master status page so that a human can watch the progress of the live computation. When aggre￾gating counter values, the master eliminates the effects of duplicate executions of the same map or reduce task to avoid double counting. (Duplicate executions can arise from our use of backup tasks and from re-execution of tasks due to failures.) Some counter values are automatically maintained by the MapReduce library, such as the number of in￾put key/value pairs processed and the number of output key/value pairs produced. Users have found the counter facility useful for san￾ity checking the behavior of MapReduce operations. For example, in some MapReduce operations, the user code may want to ensure that the number of output pairs produced exactly equals the number of input pairs pro￾cessed, or that the fraction of German documents pro￾cessed is within some tolerable fraction of the total num￾ber of documents processed. 5 Performance In this section we measure the performance of MapRe￾duce on two computations running on a large cluster of machines. One computation searches through approxi￾mately one terabyte of data looking for a particular pat￾tern. The other computation sorts approximately one ter￾abyte of data. These two programs are representative of a large sub￾set of the real programs written by users of MapReduce – one class of programs shuffles data from one representa￾tion to another, and another class extracts a small amount of interesting data from a large data set. 5.1 Cluster Configuration All of the programs were executed on a cluster that consisted of approximately 1800 machines. Each ma￾chine had two 2GHz Intel Xeon processors with Hyper￾Threading enabled, 4GB of memory, two 160GB IDE 20 40 60 80 100 Seconds 0 10000 20000 30000 Input (MB/s) Figure 2: Data transfer rate over time disks, and a gigabit Ethernet link. The machines were arranged in a two-level tree-shaped switched network with approximately 100-200 Gbps of aggregate band￾width available at the root. All of the machines were in the same hosting facility and therefore the round-trip time between any pair of machines was less than a mil￾lisecond. Out of the 4GB of memory, approximately 1-1.5GB was reserved by other tasks running on the cluster. The programs were executed on a weekend afternoon, when the CPUs, disks, and network were mostly idle. 5.2 Grep The grep program scans through 1010 100-byte records, searching for a relatively rare three-character pattern (the pattern occurs in 92,337 records). The input is split into approximately 64MB pieces (M = 15000), and the en￾tire output is placed in one file (R = 1). Figure 2 shows the progress of the computation over time. The Y-axis shows the rate at which the input data is scanned. The rate gradually picks up as more machines are assigned to this MapReduce computation, and peaks at over 30 GB/s when 1764 workers have been assigned. As the map tasks finish, the rate starts dropping and hits zero about 80 seconds into the computation. The entire computation takes approximately 150 seconds from start to finish. This includes about a minute of startup over￾head. The overhead is due to the propagation of the pro￾gram to all worker machines, and delays interacting with GFS to open the set of 1000 input files and to get the information needed for the locality optimization. 5.3 Sort The sort program sorts 1010 100-byte records (approxi￾mately 1 terabyte of data). This program is modeled after the TeraSort benchmark [10]. The sorting program consists of less than 50 lines of user code. A three-line Map function extracts a 10-byte sorting key from a text line and emits the key and the To appear in OSDI 2004 8

20000 20000 20000 15000 15000 15000 10000 10000 10000 5000 5000 5000 0 500 1000 50 1000 1000 20000- 20000 20000 15000 15000 15000 10000 10000 10000 5000 5000 5000 0 1000 500 I000 1000 20000 20000 20000 (s/)indino 15000 15000 15000 10000 0000 10000 5000 5000 5000 0 500 1000 500 1000 500 1000 Seconds Seconds Seconds (a)Normal execution (b)No backup tasks (c)200 tasks killed Figure 3:Data transfer rates over time for different executions of the sort program original text line as the intermediate key/value pair.We the first batch of approximately 1700 reduce tasks (the used a built-in Identity function as the Reduce operator. entire MapReduce was assigned about 1700 machines. This functions passes the intermediate key/value pair un- and each machine executes at most one reduce task at a changed as the output key/value pair.The final sorted time).Roughly 300 seconds into the computation,some output is written to a set of 2-way replicated GFS files of these first batch of reduce tasks finish and we start (i.e.,2 terabytes are written as the output of the program). shuffling data for the remaining reduce tasks.All of the As before,the input data is split into 64MB pieces shuffling is done about 600 seconds into the computation. (M =15000).We partition the sorted output into 4000 The bottom-left graph shows the rate at which sorted files(R=4000).The partitioning function uses the ini- data is written to the final output files by the reduce tasks. tial bytes of the key to segregate it into one of R pieces. There is a delay between the end of the first shuffling pe- Our partitioning function for this benchmark has built- riod and the start of the writing period because the ma- chines are busy sorting the intermediate data.The writes in knowledge of the distribution of keys.In a general sorting program,we would add a pre-pass MapReduce continue at a rate of about 2-4 GB/s for a while.All of the writes finish about 850 seconds into the computation. operation that would collect a sample of the keys and Including startup overhead,the entire computation takes use the distribution of the sampled keys to compute split- points for the final sorting pass. 891 seconds.This is similar to the current best reported result of 1057 seconds for the TeraSort benchmark [181. Figure 3 (a)shows the progress of a normal execution A few things to note:the input rate is higher than the of the sort program.The top-left graph shows the rate shuffle rate and the output rate because of our locality at which input is read.The rate peaks at about 13 GB/s optimization-most data is read from a local disk and and dies off fairly quickly since all map tasks finish be- bypasses our relatively bandwidth constrained network. fore 200 seconds have elapsed.Note that the input rate The shuffle rate is higher than the output rate because is less than for grep.This is because the sort map tasks the output phase writes two copies of the sorted data (we spend about half their time and I/O bandwidth writing in- make two replicas of the output for reliability and avail- termediate output to their local disks.The corresponding ability reasons).We write two replicas because that is intermediate output for grep had negligible size. the mechanism for reliability and availability provided The middle-left graph shows the rate at which data by our underlying file system.Network bandwidth re- is sent over the network from the map tasks to the re- quirements for writing data would be reduced if the un- duce tasks.This shuffling starts as soon as the first derlying file system used erasure coding [14]rather than map task completes.The first hump in the graph is for replication. To appear in OSDI 2004 9

500 1000 0 5000 10000 15000 20000 Input (MB/s) 500 1000 0 5000 10000 15000 20000 Shuffle (MB/s) 500 1000 Seconds 0 5000 10000 15000 20000 Output (MB/s) Done (a) Normal execution 500 1000 0 5000 10000 15000 20000 Input (MB/s) 500 1000 0 5000 10000 15000 20000 Shuffle (MB/s) 500 1000 Seconds 0 5000 10000 15000 20000 Output (MB/s) Done (b) No backup tasks 500 1000 0 5000 10000 15000 20000 Input (MB/s) 500 1000 0 5000 10000 15000 20000 Shuffle (MB/s) 500 1000 Seconds 0 5000 10000 15000 20000 Output (MB/s) Done (c) 200 tasks killed Figure 3: Data transfer rates over time for different executions of the sort program original text line as the intermediate key/value pair. We used a built-in Identity function as the Reduce operator. This functions passes the intermediate key/value pair un￾changed as the output key/value pair. The final sorted output is written to a set of 2-way replicated GFS files (i.e., 2 terabytes are written as the output of the program). As before, the input data is split into 64MB pieces (M = 15000). We partition the sorted output into 4000 files (R = 4000). The partitioning function uses the ini￾tial bytes of the key to segregate it into one of R pieces. Our partitioning function for this benchmark has built￾in knowledge of the distribution of keys. In a general sorting program, we would add a pre-pass MapReduce operation that would collect a sample of the keys and use the distribution of the sampled keys to compute split￾points for the final sorting pass. Figure 3 (a) shows the progress of a normal execution of the sort program. The top-left graph shows the rate at which input is read. The rate peaks at about 13 GB/s and dies off fairly quickly since all map tasks finish be￾fore 200 seconds have elapsed. Note that the input rate is less than for grep. This is because the sort map tasks spend about half their time and I/O bandwidth writing in￾termediate output to their local disks. The corresponding intermediate output for grep had negligible size. The middle-left graph shows the rate at which data is sent over the network from the map tasks to the re￾duce tasks. This shuffling starts as soon as the first map task completes. The first hump in the graph is for the first batch of approximately 1700 reduce tasks (the entire MapReduce was assigned about 1700 machines, and each machine executes at most one reduce task at a time). Roughly 300 seconds into the computation, some of these first batch of reduce tasks finish and we start shuffling data for the remaining reduce tasks. All of the shuffling is done about 600 secondsinto the computation. The bottom-left graph shows the rate at which sorted data is written to the final output files by the reduce tasks. There is a delay between the end of the first shuffling pe￾riod and the start of the writing period because the ma￾chines are busy sorting the intermediate data. The writes continue at a rate of about 2-4 GB/s for a while. All of the writes finish about 850 seconds into the computation. Including startup overhead, the entire computation takes 891 seconds. This is similar to the current best reported result of 1057 seconds for the TeraSort benchmark [18]. A few things to note: the input rate is higher than the shuffle rate and the output rate because of our locality optimization – most data is read from a local disk and bypasses our relatively bandwidth constrained network. The shuffle rate is higher than the output rate because the output phase writes two copies of the sorted data (we make two replicas of the output for reliability and avail￾ability reasons). We write two replicas because that is the mechanism for reliability and availability provided by our underlying file system. Network bandwidth re￾quirements for writing data would be reduced if the un￾derlying file system used erasure coding [14] rather than replication. To appear in OSDI 2004 9

5.4 Effect of Backup Tasks 1000 In Figure 3(b),we show an execution of the sort pro- 800 gram with backup tasks disabled.The execution flow is similar to that shown in Figure 3 (a),except that there is 600 a very long tail where hardly any write activity occurs. After 960 seconds,all except 5 of the reduce tasks are 400 completed.However these last few stragglers don't fin- ish until 300 seconds later.The entire computation takes 1283 seconds,an increase of 44%in elapsed time. qu 200 5.5 Machine Failures 2003/09 200403 2004006 2004009 In Figure 3(c),we show an execution of the sort program 3/06 200312 where we intentionally killed 200 out of 1746 worker processes several minutes into the computation.The Figure 4:MapReduce instances over time underlying cluster scheduler immediately restarted new worker processes on these machines (since only the pro- Number of jobs 29.423 cesses were killed,the machines were still functioning Average job completion time 634 secs properly). Machine days used 79,186days Input data read 3.288TB The worker deaths show up as a negative input rate Intermediate data produced 758TB since some previously completed map work disappears Output data written 193TB (since the corresponding map workers were killed)and Average worker machines per job 157 needs to be redone.The re-execution of this map work Average worker deaths per job 1.2 happens relatively quickly.The entire computation fin- Average map tasks per job 3.351 ishes in 933 seconds including startup overhead (just an Average reduce tasks per job 55 increase of 5%over the normal execution time). Unique map implementations 395 Unique reduce implementations 269 Unique map/reduce combinations 426 Experience Table 1:MapReduce jobs run in August 2004 We wrote the first version of the MapReduce library in February of 2003,and made significant enhancements to it in August of 2003,including the locality optimization, Figure 4 shows the significant growth in the number of dynamic load balancing of task execution across worker separate MapReduce programs checked into our primary machines,etc.Since that time,we have been pleasantly source code management system over time,from 0 in surprised at how broadly applicable the MapReduce li- early 2003 to almost 900 separate instances as of late brary has been for the kinds of problems we work on. September 2004.MapReduce has been so successful be- It has been used across a wide range of domains within cause it makes it possible to write a simple program and Google,including: run it efficiently on a thousand machines in the course of half an hour,greatly speeding up the development and large-scale machine learning problems, prototyping cycle.Furthermore,it allows programmers who have no experience with distributed and/or parallel .clustering problems for the Google News and systems to exploit large amounts of resources easily. Froogle products, At the end of each job.the MapReduce library logs .extraction of data used to produce reports of popular statistics about the computational resources used by the queries (e.g.Google Zeitgeist), job.In Table 1,we show some statistics for a subset of MapReduce jobs run at Google in August 2004. extraction of properties of web pages for new exper- iments and products (e.g.extraction of geographi- cal locations from a large corpus of web pages for 6.1 Large-Scale Indexing localized search),and One of our most significant uses of MapReduce to date large-scale graph computations has been a complete rewrite of the production index- To appear in OSDI 2004 10

5.4 Effect of Backup Tasks In Figure 3 (b), we show an execution of the sort pro￾gram with backup tasks disabled. The execution flow is similar to that shown in Figure 3 (a), except that there is a very long tail where hardly any write activity occurs. After 960 seconds, all except 5 of the reduce tasks are completed. However these last few stragglers don’t fin￾ish until 300 seconds later. The entire computation takes 1283 seconds, an increase of 44% in elapsed time. 5.5 Machine Failures In Figure 3 (c), we show an execution of the sort program where we intentionally killed 200 out of 1746 worker processes several minutes into the computation. The underlying cluster scheduler immediately restarted new worker processes on these machines (since only the pro￾cesses were killed, the machines were still functioning properly). The worker deaths show up as a negative input rate since some previously completed map work disappears (since the corresponding map workers were killed) and needs to be redone. The re-execution of this map work happens relatively quickly. The entire computation fin￾ishes in 933 seconds including startup overhead (just an increase of 5% over the normal execution time). 6 Experience We wrote the first version of the MapReduce library in February of 2003, and made significant enhancements to it in August of 2003, including the locality optimization, dynamic load balancing of task execution across worker machines, etc. Since that time, we have been pleasantly surprised at how broadly applicable the MapReduce li￾brary has been for the kinds of problems we work on. It has been used across a wide range of domains within Google, including: • large-scale machine learning problems, • clustering problems for the Google News and Froogle products, • extraction of data used to produce reports of popular queries (e.g. Google Zeitgeist), • extraction of properties of web pages for new exper￾iments and products (e.g. extraction of geographi￾cal locations from a large corpus of web pages for localized search), and • large-scale graph computations. 2003/03 2003/06 2003/09 2003/12 2004/03 2004/06 2004/09 0 200 400 600 800 1000 Number of instances in source tree Figure 4: MapReduce instances over time Number of jobs 29,423 Average job completion time 634 secs Machine days used 79,186 days Input data read 3,288 TB Intermediate data produced 758 TB Output data written 193 TB Average worker machines per job 157 Average worker deaths per job 1.2 Average map tasks per job 3,351 Average reduce tasks per job 55 Unique map implementations 395 Unique reduce implementations 269 Unique map/reduce combinations 426 Table 1: MapReduce jobs run in August 2004 Figure 4 shows the significant growth in the number of separate MapReduce programs checked into our primary source code management system over time, from 0 in early 2003 to almost 900 separate instances as of late September 2004. MapReduce has been so successful be￾cause it makes it possible to write a simple program and run it efficiently on a thousand machines in the course of half an hour, greatly speeding up the development and prototyping cycle. Furthermore, it allows programmers who have no experience with distributed and/or parallel systems to exploit large amounts of resources easily. At the end of each job, the MapReduce library logs statistics about the computational resources used by the job. In Table 1, we show some statistics for a subset of MapReduce jobs run at Google in August 2004. 6.1 Large-Scale Indexing One of our most significant uses of MapReduce to date has been a complete rewrite of the production index￾To appear in OSDI 2004 10