VIEWS: 0 PAGES: 39 POSTED ON: 12/18/2013
MapReduce Outline r MapReduce overview r Note: These notes are based on notes provided by Google What is a Cloud? q Cloud = Lots of storage + compute cycles nearby Data-Intensive Computing q Data-Intensive m Typically store data at datacenters m Use compute nodes nearby m Compute nodes run computation services r In data-intensive computing, the focus is on the data: problem areas include m Storage m Communication bottleneck m Moving tasks to data (rather than vice-versa) m Security m Availability of Data m Scalability Computation Services q Google → MapReduce, Sawzall r Yahoo → Hadoop, Pig Latin r Microsoft → Dryad, DryadLINQ Motivation: Large Scale Data Processing r Want to process lots of data ( > 1 TB) r Want to parallelize across hundreds/thousands of CPUs m How to parallelize m How to distribute m How to handle failures r Want to make this easy What is MapReduce? r MapReduce is an abstraction that allows programmers to specify computations that can be done in parallel r MapReduce hides the messy details needed to support the computations e.g., m Distribution and synchronization m Machine failures m Data distribution m Load balancing r This is widely used at Google Programming Model r MapReduce simplifies programming through its library. r The user of the MapReduce library expresses the computation as two functions: Map, Reduce Programming Model r Map m Takes an input pair and produces a set of intermediate key/value pairs e.g., • Map: (key1, value1) à list(key2,value2) r The MapReduce library groups together all intermediate values associated with the same intermediate key r Reduce m This function accepts an intermediate key and a set of values for that key m Reduce: (key2,list(key2,value2)) à value3 Example: Word Frequencies in Web Pages r Determine the count of each word that appears in a document (or a set of documents) m Each file is associated with a document URL r Map function m Key = document URL m Value = document contents r Output of map function is (potentially many) key/value pairs m Output (word, “1”) once per word in the document Example: Word Frequencies in Web Pages r Pseudo code for map Map(String key, String value): // input_key: document name // input_value: document contents for each word w in value: EmitIntermediate(w, "1"); Example: Word Frequencies in Web Pages r Example key, value pair: m “document_example”, “to be or not to be” r Result of applying the map function m “to”, 1 m “be”, 1 m “or”, 1 m “not”, 1 m “to”, 1 m “be”, 1 Example: Word Frequencies in Web Pages r Pseudo-code for Reduce Reduce(String key, values): // key: a word, same for input and output // values: a list of counts int result = 0; for each v in values: result = result + value; Emit(result); The function sums together all counts emitted for a particular word Example: Word Frequencies in Web Pages r The MapReduce framework sorts all pairs with the same key m (be,1), (be,1), (not,1), (or, 1), (to, 1), (to,1) r The pairs are then grouped m (be, 1,1), (not, 1), (or, 1), (to, 1, 1) r The reduce function combines (sums) the values for a key m Example: Applying reduce to (be, 1, 1) = 2 Example: Distributed Grep r Find all occurrences of a given pattern in a a file (or set of files) r Input consists of (url+offset, line) r map(key=url+offset, val=line): m If contents match specified pattern, emit (line, “1) r reduce(key=line, values=uniq_counts): m Example of input to reduce is essentially (line, [1,1,1,1]) m Don’t do anything; just emit line Example: Count of URL Access Frequency r Map function m Input: <log of web page requests, content of log> m Outputs: <URL, 1> r Reduce function adds together all values for the same URL Example:Web structure r Simple representation of WWW link graph m Map • Input: (URL, page-contents) • Output: (URL, list-of-URLs) r Who maps to me? m Map • Input: (URL, list-of-URLS) • Output: For each u in list-of-URLS output <u,URL> m Reduce: Concatenates the list of all source URLs associated with u and emits (<u, list(URL)) The Infrastructure r Large clusters of commodity PCs and networking hardware r Clusters consists of 100/1000s of machines (failures are common) r GFS (Google File System). m Distributed file system. m Provides replication of the data. The Infrastructure r Users submit jobs to a scheduling system r Possible partitions of data can be based on files, databases, file lines, database records etc; Execution r Map invocations are distributed across multiple machines by automatically partitioning the input data into a set of M splits. r The input splits can be processed in parallel by different machines r Reduce invocations are distributed by partitioning the intermediate key space into R pieces using a hash function: hash(key) mod R m R and the partitioning function are specified by the programmer. Execution Workers are assigned work by the master The master is started by the MapReduce Framework Execution Workers assigned map tasks read the input, parse it and invoke the user’s Map() method. Execution • Intermediate key/value pairs are buffered in memory •Periodically, buffered data is written to local disk (R files) •Pseudo random partitioning function (e.g., (hash(k) mod R) Execution •Locations are passed back to the master who forwards these locations to workers executing the reduce function. Execution • Reduce runs after all mappers are done • Workers executing Reduce are notified by the master about location of intermediate data Execution • Reduce workers use remote procedure calls to read the data from local disks of map works • Sorts all intermediate data by intermediate key Execution • Reduce worker iterates over the sorted intermediate data and for each key encountered it passes the key and the corresponding set of intermediate values to the Reduce function Execution • The output of the Reduce function is appended to a final output file Data flow r Input, final output are stored on a distributed file system m Scheduler tries to schedule map tasks “close” to physical storage location of input data r Intermediate results are stored on local file system of map and reduce workers r Output can be input to another map reduce task Execution Parallel Execution Coordination r Master data structures m Task status: (idle, in-progress, completed) m Idle tasks get scheduled as workers become available m When a map task completes, it sends the master the location and sizes of its R intermediate files, one for each reducer m Master pushes this info to reducers r Master pings workers periodically to detect failures Failures r Map worker failure m Map tasks completed or in-progress at worker are reset to idle m Reduce workers are notified when task is rescheduled on another worker r Reduce worker failure m Only in-progress tasks are reset to idle r Master failure m MapReduce task is aborted and client is notified Locality r MapReduce master takes the location information of input files into account and attempts to schedule a map task on a machine that contains a replica of the corresponding input data r Schedule a map task near a replica of that task’s input data r The goal is to read most input data locally and thus reduce the consumption of network bandwidth Task Granularity r M and R should be much larger than the number of available machines. m Dynamic load balancing. m Speeds up recovery in case of failures. r R determines the number of output files m Often constrained by users. Backup Tasks r Stragglers - A common reason for long computations. r Schedule backups for remaining jobs (in progress jobs) when map or reduce phases near completion. m Slightly increases needed computational resources. m Does not increase running time, but has the potential to improve it significantly. Combiners r Often a map task will produce many pairs of the form (k,v1), (k,v2), … for the same key k m E.g., popular words in Word Count r Can save network time by pre-aggregating at mapper m combine(k1, list(v1)) à v2 m Usually same as reduce function r Works only if reduce function is commutative and associative Partition Function r Inputs to map tasks are created by contiguous splits of input file r For reduce, we need to ensure that records with the same intermediate key end up at the same worker r System uses a default partition function e.g., hash(key) mod R r Sometimes useful to override; What if all output keys are URLS and we want all entries for a single host to end up in the same output file? m Use hash(hostname(URL)) mod R ensures URLs from a host end up in the same output file Summary r MapReduce – a framework for distributed computing. m Distributed programs are easy to write and understand. m Provides fault tolerance m Program execution can be easily monitored. r It works for Google!!
Pages to are hidden for
"MapReduce"Please download to view full document