MapReduce: Simplified Data Processing on Large Clusters PDF

MapReduce: Simplified Data Processing on Large Clusters Jeffrey Dean and Sanjay Ghemawat [email protected], [email protected] Google, Inc. Abst...

MapReduce: Simplified Data Processing on Large Clusters Jeffrey Dean and Sanjay Ghemawat [email protected], [email protected] 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 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 of string inputs and outputs, conceptually the map and 2 Programming Model 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) → list(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 matches a supplied pattern. The reduce function is an Consider the problem of counting the number of oc- 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): tion processes logs of web page requests and outputs // key: document name // value: document contents hURL, 1i. The reduce function adds together all values for each word w in value: for the same URL and emits a hURL, total counti 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 htarget, sourcei 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)); htarget, list(source)i 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 hword, f requencyi pairs. The In addition, the user writes code to fill in a mapreduce map function emits a hhostname, term vectori 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. 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 file 0 (5) remote read split 2 (3) read (4) local write worker output split 3 worker 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 hword, document IDi switched Ethernet. 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. hword, list(document ID)i 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 hkey, recordi 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 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- consists of a set of tasks, and is mapped by the scheduler terface are possible. The right choice depends on the 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 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 master. The rest are workers that are assigned work (for non-idle tasks). 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 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 R1 is equivalent to maintenance on a running cluster was causing groups of the output for R1 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 R1 and R2. Let e(Ri ) be the execution of Ri that committed (there is exactly one such execution). The weaker semantics Master Failure arise because e(R1 ) 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 ) 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 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 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 tricky, since the actual computation happens in a dis- A reader does not necessarily need to provide data 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- The master runs an internal HTTP server and exports rary file and atomically renames this file once it has been a set of status pages for human consumption. The sta- fully generated. 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 7  Counter* uppercase; 30000 Input (MB/s) uppercase = GetCounter("uppercase"); 20000 map(String name, String contents): for each word w in contents: 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 at over 30 GB/s when 1764 workers have been assigned. In this section we measure the performance of MapRe- 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- computation takes approximately 150 seconds from start mately one terabyte of data looking for a particular pat- 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 The sort program sorts 1010 100-byte records (approxi- 5.1 Cluster Configuration mately 1 terabyte of data). This program is modeled after All of the programs were executed on a cluster that the TeraSort benchmark. 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 8  20000 20000 20000 Input (MB/s) Done Done Done Input (MB/s) 15000 15000 Input (MB/s) 15000 10000 10000 10000 5000 5000 5000 0 0 0 500 1000 500 1000 500 1000 20000 20000 20000 Shuffle (MB/s) Shuffle (MB/s) Shuffle (MB/s) 15000 15000 15000 10000 10000 10000 5000 5000 5000 0 0 0 500 1000 500 1000 500 1000 20000 20000 20000 Output (MB/s) Output (MB/s) Output (MB/s) 15000 15000 15000 10000 10000 10000 5000 5000 5000 0 0 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- riod and the start of the writing period because the ma- Our partitioning function for this benchmark has built- chines are busy sorting the intermediate data. The writes in knowledge of the distribution of keys. In a general continue at a rate of about 2-4 GB/s for a while. All of sorting program, we would add a pre-pass MapReduce 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- 891 seconds. This is similar to the current best reported points for the final sorting pass. result of 1057 seconds for the TeraSort benchmark. 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 rather than map task completes. The first hump in the graph is for replication. To appear in OSDI 2004 9  1000 5.4 Effect of Backup Tasks Number of instances in source tree 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 200 1283 seconds, an increase of 44% in elapsed time. 0 5.5 Machine Failures 2003/03 2003/06 2003/09 2003/12 2004/03 2004/06 2004/09 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 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 79,186 days Input data read 3,288 TB The worker deaths show up as a negative input rate Intermediate data produced 758 TB since some previously completed map work disappears Output data written 193 TB (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 6 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 ing system that produces the data structures used for the make it easier for programmers to write parallel pro- Google web search service. The indexing system takes grams. A key difference between these systems and as input a large set of documents that have been retrieved MapReduce is that MapReduce exploits a restricted pro- by our crawling system, stored as a set of GFS files. The gramming model to parallelize the user program auto- raw contents for these documents are more than 20 ter- matically and to provide transparent fault-tolerance. abytes of data. The indexing process runs as a sequence Our locality optimization draws its inspiration from of five to ten MapReduce operations. Using MapReduce techniques such as active disks [12, 15], where compu- (instead of the ad-hoc distributed passes in the prior ver- tation is pushed into processing elements that are close sion of the indexing system) has provided several bene- to local disks, to reduce the amount of data sent across fits: I/O subsystems or the network. We run on commodity processors to which a small number of disks are directly The indexing code is simpler, smaller, and easier to connected instead of running directly on disk controller understand, because the code that deals with fault processors, but the general approach is similar. tolerance, distribution and parallelization is hidden Our backup task mechanism is similar to the eager within the MapReduce library. For example, the scheduling mechanism employed in the Charlotte Sys- size of one phase of the computation dropped from tem. One of the shortcomings of simple eager approximately 3800 lines of C++ code to approx- scheduling is that if a given task causes repeated failures, imately 700 lines when expressed using MapRe- the entire computation fails to complete. We fix some in- duce. stances of this problem with our mechanism for skipping The performance of the MapReduce library is good bad records. enough that we can keep conceptually unrelated The MapReduce implementation relies on an in-house computations separate, instead of mixing them to- cluster management system that is responsible for dis- gether to avoid extra passes over the data. This tributing and running user tasks on a large collection of makes it easy to change the indexing process. For shared machines. Though not the focus of this paper, the example, one change that took a few months to cluster management system is similar in spirit to other make in our old indexing system took only a few systems such as Condor. days to implement in the new system. The sorting facility that is a part of the MapReduce library is similar in operation to NOW-Sort. Source The indexing process has become much easier to machines (map workers) partition the data to be sorted operate, because most of the problems caused by and send it to one of R reduce workers. Each reduce machine failures, slow machines, and networking worker sorts its data locally (in memory if possible). Of hiccups are dealt with automatically by the MapRe- course NOW-Sort does not have the user-definable Map duce library without operator intervention. Further- and Reduce functions that make our library widely appli- more, it is easy to improve the performance of the cable. indexing process by adding new machines to the in- River provides a programming model where pro- dexing cluster. cesses communicate with each other by sending data over distributed queues. Like MapReduce, the River 7 Related Work system tries to provide good average case performance even in the presence of non-uniformities introduced by Many systems have provided restricted programming heterogeneous hardware or system perturbations. River models and used the restrictions to parallelize the com- achieves this by careful scheduling of disk and network putation automatically. For example, an associative func- transfers to achieve balanced completion times. MapRe- tion can be computed over all prefixes of an N element duce has a different approach. By restricting the pro- array in log N time on N processors using parallel prefix gramming model, the MapReduce framework is able computations [6, 9, 13]. MapReduce can be considered to partition the problem into a large number of fine- a simplification and distillation of some of these models grained tasks. These tasks are dynamically scheduled based on our experience with large real-world compu- on available workers so that faster workers process more tations. More significantly, we provide a fault-tolerant tasks. The restricted programming model also allows implementation that scales to thousands of processors. us to schedule redundant executions of tasks near the In contrast, most of the parallel processing systems have end of the job which greatly reduces completion time in only been implemented on smaller scales and leave the the presence of non-uniformities (such as slow or stuck details of handling machine failures to the programmer. workers). Bulk Synchronous Programming and some MPI BAD-FS has a very different programming model primitives provide higher-level abstractions that from MapReduce, and unlike MapReduce, is targeted to To appear in OSDI 2004 11 the execution of jobs across a wide-area network. How- David Kramer, Shun-Tak Leung, and Josh Redstone for ever, there are two fundamental similarities. (1) Both their work in developing GFS. We would also like to systems use redundant execution to recover from data thank Percy Liang and Olcan Sercinoglu for their work loss caused by failures. (2) Both use locality-aware in developing the cluster management system used by scheduling to reduce the amount of data sent across con- MapReduce. Mike Burrows, Wilson Hsieh, Josh Leven- gested network links. berg, Sharon Perl, Rob Pike, and Debby Wallach pro- TACC is a system designed to simplify con- vided helpful comments on earlier drafts of this pa- struction of highly-available networked services. Like per. The anonymous OSDI reviewers, and our shepherd, MapReduce, it relies on re-execution as a mechanism for Eric Brewer, provided many useful suggestions of areas implementing fault-tolerance. where the paper could be improved. Finally, we thank all the users of MapReduce within Google’s engineering or- ganization for providing helpful feedback, suggestions, 8 Conclusions and bug reports. The MapReduce programming model has been success- fully used at Google for many different purposes. We References attribute this success to several reasons. First, the model is easy to use, even for programmers without experience Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau, with parallel and distributed systems, since it hides the David E. Culler, Joseph M. Hellerstein, and David A. Pat- details of parallelization, fault-tolerance, locality opti- terson. High-performance sorting on networks of work- mization, and load balancing. Second, a large variety stations. In Proceedings of the 1997 ACM SIGMOD In- ternational Conference on Management of Data, Tucson, of problems are easily expressible as MapReduce com- Arizona, May 1997. putations. For example, MapReduce is used for the gen- eration of data for Google’s production web search ser- Remzi H. Arpaci-Dusseau, Eric Anderson, Noah vice, for sorting, for data mining, for machine learning, Treuhaft, David E. Culler, Joseph M. Hellerstein, David and many other systems. Third, we have developed an Patterson, and Kathy Yelick. Cluster I/O with River: Making the fast case common. In Proceedings of the Sixth implementation of MapReduce that scales to large clus- Workshop on Input/Output in Parallel and Distributed ters of machines comprising thousands of machines. The Systems (IOPADS ’99), pages 10–22, Atlanta, Georgia, implementation makes efficient use of these machine re- May 1999. sources and therefore is suitable for use on many of the large computational problems encountered at Google. Arash Baratloo, Mehmet Karaul, Zvi Kedem, and Peter Wyckoff. Charlotte: Metacomputing on the web. In Pro- We have learned several things from this work. First, ceedings of the 9th International Conference on Parallel restricting the programming model makes it easy to par- and Distributed Computing Systems, 1996. allelize and distribute computations and to make such computations fault-tolerant. Second, network bandwidth Luiz A. Barroso, Jeffrey Dean, and Urs Hölzle. Web search for a planet: The Google cluster architecture. IEEE is a scarce resource. A number of optimizations in our Micro, 23(2):22–28, April 2003. system are therefore targeted at reducing the amount of data sent across the network: the locality optimization al- John Bent, Douglas Thain, Andrea C.Arpaci-Dusseau, lows us to read data from local disks, and writing a single Remzi H. Arpaci-Dusseau, and Miron Livny. Explicit copy of the intermediate data to local disk saves network control in a batch-aware distributed file system. In Pro- ceedings of the 1st USENIX Symposium on Networked bandwidth. Third, redundant execution can be used to Systems Design and Implementation NSDI, March 2004. reduce the impact of slow machines, and to handle ma- chine failures and data loss. Guy E. Blelloch. Scans as primitive parallel operations. IEEE Transactions on Computers, C-38(11), November 1989. Acknowledgements Armando Fox, Steven D. Gribble, Yatin Chawathe, Eric A. Brewer, and Paul Gauthier. Cluster-based scal- Josh Levenberg has been instrumental in revising and able network services. In Proceedings of the 16th ACM extending the user-level MapReduce API with a num- Symposium on Operating System Principles, pages 78– ber of new features based on his experience with using 91, Saint-Malo, France, 1997. MapReduce and other people’s suggestions for enhance- Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Le- ments. MapReduce reads its input from and writes its ung. The Google file system. In 19th Symposium on Op- output to the Google File System. We would like to erating Systems Principles, pages 29–43, Lake George, thank Mohit Aron, Howard Gobioff, Markus Gutschke, New York, 2003. To appear in OSDI 2004 12  S. Gorlatch. Systematic efficient parallelization of scan if (start < i) and other list homomorphisms. In L. Bouge, P. Fraigni- Emit(text.substr(start,i-start),"1"); aud, A. Mignotte, and Y. Robert, editors, Euro-Par’96. } } Parallel Processing, Lecture Notes in Computer Science }; 1124, pages 401–408. Springer-Verlag, 1996. REGISTER_MAPPER(WordCounter); Jim Gray. Sort benchmark home page. http://research.microsoft.com/barc/SortBenchmark/. // User’s reduce function class Adder : public Reducer { William Gropp, Ewing Lusk, and Anthony Skjellum. virtual void Reduce(ReduceInput* input) { Using MPI: Portable Parallel Programming with the // Iterate over all entries with the Message-Passing Interface. MIT Press, Cambridge, MA, // same key and add the values int64 value = 0; 1999. while (!input->done()) { L. Huston, R. Sukthankar, R. Wickremesinghe, M. Satya- value += StringToInt(input->value()); narayanan, G. R. Ganger, E. Riedel, and A. Ailamaki. Di- input->NextValue(); amond: A storage architecture for early discard in inter- } active search. In Proceedings of the 2004 USENIX File // Emit sum for input->key() and Storage Technologies FAST Conference, April 2004. Emit(IntToString(value)); Richard E. Ladner and Michael J. Fischer. Parallel prefix } computation. Journal of the ACM, 27(4):831–838, 1980. }; REGISTER_REDUCER(Adder); Michael O. Rabin. Efficient dispersal of information for security, load balancing and fault tolerance. Journal of int main(int argc, char** argv) { the ACM, 36(2):335–348, 1989. ParseCommandLineFlags(argc, argv); Erik Riedel, Christos Faloutsos, Garth A. Gibson, and MapReduceSpecification spec; David Nagle. Active disks for large-scale data process- ing. IEEE Computer, pages 68–74, June 2001. // Store list of input files into "spec" for (int i = 1; i < argc; i++) { Douglas Thain, Todd Tannenbaum, and Miron Livny. MapReduceInput* input = spec.add_input(); Distributed computing in practice: The Condor experi- input->set_format("text"); ence. Concurrency and Computation: Practice and Ex- input->set_filepattern(argv[i]); perience, 2004. input->set_mapper_class("WordCounter"); } L. G. Valiant. A bridging model for parallel computation. Communications of the ACM, 33(8):103–111, 1997. // Specify the output files: // /gfs/test/freq-00000-of-00100 Jim Wyllie. Spsort: How to sort a terabyte quickly. // /gfs/test/freq-00001-of-00100 http://alme1.almaden.ibm.com/cs/spsort.pdf. //... MapReduceOutput* out = spec.output(); out->set_filebase("/gfs/test/freq"); A Word Frequency out->set_num_tasks(100); out->set_format("text"); out->set_reducer_class("Adder"); This section contains a program that counts the number of occurrences of each unique word in a set of input files // Optional: do partial sums within map specified on the command line. // tasks to save network bandwidth out->set_combiner_class("Adder"); #include "mapreduce/mapreduce.h" // Tuning parameters: use at most 2000 // User’s map function // machines and 100 MB of memory per task class WordCounter : public Mapper { spec.set_machines(2000); public: spec.set_map_megabytes(100); virtual void Map(const MapInput& input) { spec.set_reduce_megabytes(100); const string& text = input.value(); const int n = text.size(); // Now run it for (int i = 0; i < n; ) { MapReduceResult result; // Skip past leading whitespace if (!MapReduce(spec, &result)) abort(); while ((i < n) && isspace(text[i])) i++; // Done: ’result’ structure contains info // about counters, time taken, number of // Find word end // machines used, etc. int start = i; while ((i < n) && !isspace(text[i])) return 0; i++; } To appear in OSDI 2004 13