Large Scale Text Analysis Using the Map/Reduce

Transcription

1 Large Scale Text Analysis Using the Map/Reduce Hierarchy David Buttler This work is performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 Lawrence Livermore National Laboratory

2 Large scale computing with commodity hardware Origins Google GFS, Map/Reduce, BigTable Microsoft Azure Hadoop: Yahoo! / Open Source Software Why do we care: k-mer lexing 10 hours on a single fat node 1 hour on an old cluster 2

3 The M/R stack of open source software Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 3

4 The M/R stack of open source software HDFS Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 4

5 HDFS Replicated Distributed ib t Centrally managed Data Node 1 SPOF Data a Node 2 Limited number of files Not POSIX compliant Rack-aware Name Node File: path, {Blocks} Data Node n File: path, {Blocks} 5

6 The M/R stack of open source software Map / Reduce Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 6

7 Map/Reduce is functional programming distributed over a cluster Distributed computation Two phase computation ti Built-in shuffle/sort between phases Canonical example: word frequency count for the web Map to, 1 be, 1 or, 1 not, 1 to, 1 be, 1 Input Documents shuffle / sort be, 1 be, 1 not, 1 Reduce to, 1 to, 1 be, 2+ not, 1+ to, 2+ 7

8 More interesting M/R examples Map input: document Map output: t raw text t Map input: text Map output: Named entity annotations 8

9 The M/R stack of open source software Zookeeper Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 9

10 Zookeeper A highly available, scalable, distributed, configuration, consensus, group membership, leader election, naming, and coordination service Uses: HBase: row locking; region key ranges; region server addresses Katta: shard location information Message queues Not: a large scale data store 10

11 Zookeeper Guarantees 1. Clients will never detect old data. 2. Clients will get notified of a change to data they are watching within a bounded period of time. 3. All requests from a client will be processed in order. 4. All results received by a client will be consistent with results received by all other clients. 11

12 Zookeeper Data Model Hierarchal namespace Each znode has data and children data is read and written in its entirety Nodes store < 1MB data Writes go to all nodes / services YaView servers locks Name 1 Name n read-1 HBase Katta 12

13 ZooKeeper Service ZooKeeper Service Leader Server Server Server Server Server Client Client Client Client Client Client Client All servers store a copy of the data (in memory) A leader is elected at startup Followers service clients, all updates go through leader Update responses are sent when a majority of servers have persisted the change 13

14 The M/R stack of open source software HBase Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 14

15 HBase Distributed column oriented data store Only supports one data type Tables are broken into regions Regions are automatically split and redistributed All data is local Scales to > 1M row / second insert rate (20 node cluster) Tightly integrated with Hadoop -> rows can be input/output t t for map/reduce tasks 15

16 HBase Data model Table Row ID Column Family Map Family Qualifier Map Qualifier Version Value Physical files 16

17 HBase Data model (simplified) Table Row ID Key Family: Column: Version Value aue Regions partitioned i on row key 17

18 HBase System Architecture From 18

19 HBase Master manages region servers 19

20 Hbase Client directly access region servers for data 20

21 The M/R stack of open source software Hive Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 21

22 Hive provides and SQL-like interface to data Components Shell: SQL-like command line; Web; JDBC Driver: API interface Compiler: parse, plan, optimize Execution Engine: DAG of stages (M/R, HDFS, or metadata) Metastore: schema, location in HDFS, SerDe 22

23 The M/R stack of open source software Solr / Katta Workflow (Cascading / Azkaban) Katta Solr / Lucene Pig Hive Zookeeper HBase Map / Reduce HDFS 23

24 Solr Faceted text search interface built on top of Lucene Built as a native web app drops into any web server 24

25 Faceted search is a foundational component for ad hoc document analysis 25

26 Solr architecture HTTP Request Servlet Update Servlet Admin Interface Standard Request Handler Disjunction Max Request Handler Custom XML XML Request Response Update Handler Writer Interface Config Schema Caching Analysis Solr Core Concurrency Update Handler Replication Lucene Diagram by Yonik Seeley 26

27 Katta provides vertical and horizontal scalability Shard A Node 1 Shard Z 1) Query 5) Combined Results Text search interface Solr Node n Shard A Shard Z Zookeeper 27

28 Projects using Hadoop at LLNL Student projects Bioinformatics i [James Leek] Continuous time LDA [Kurt Miller and Tina Elliasi-Rad] Advanced R&D projects Network analytics Keyword tagging g and entity extraction Faceted Search Research projects READ LDRD Program deployments BKMS 28

29 Bioinformatics (student project) KPATH: produce DNA signatures for detection of pathogens k-mer lexing: produce set of unique DNA sequences of length Sliding window Discover k-mers that are unique between bacteria and viruses 29

30 K-mer parsing performance comparisons Lexing bacteria file 30 k-mer length [120 GB] Optimized i suffix tree [ C implementation] ti on single node, 256 GB RAM, 16 processor system 10.5 hours Custom hadoop implementation 85 nodes, 8 GB RAM, dual processor [old] ~1 hour 30

31 Unique K-mer grouping performance Group unique k-mers of length 15 [13 GB data] Pig implementation ti using outer joins [10 LOC] More than 9 hours Customer hadoop implementation: 3 hours 26 minutes 31

32 Network data HDFS provides storage layer for large repositories of network data Hive provides an SQL interface Performance on single query for 6 months of data: Tuned Oracle DB: hours to days Hive: minutes 32

33 Hadoop-based document management architecture HBase RSS Pub Med Ingest cumen nt Source Text DoMetadata Annotations Process Katta/Solr Pubmed NYT Access Access Tomcat Document Viewer RSS etc. Faceted Search 33

34 Example Data Flow Initial Load Parser NLP Topics Index Load original documents into document table Custom map code to extract text and meta data Named Entity Extraction (SNER) Parsing / Coreference Send corpus slices to LDA for topic modeling Write specific HBase columns to faceted Solr index shards Serve Manage indexes with Katta over HDFS 34

35 Keyword tagging & Entity extraction Keyword tagging Large dictionaries i (100K terms) Finite state machine to store dictionary Named Entity Recognition Stanford NER CRF model [People, Organizations, Location] 35

36 Performance of Keyword tagging & Entity extraction 21M Pubmed entries + 1M news articles 11M Pubmed abstracts bt t 55K dictionary key phrases 6 node cluster [16 core, 96 GB RAM, 6 TB disk] Keyword Tagging g 8 minutes, 34 seconds Named Entity Annotation 1 hr 58 minutes 36

37 37

38 38

39 39

40 Faceted Search Indexing Performance Creating 1 Solr index on 1M news articles: 8hrs 16 min Map: 37 min Reduce: 8 hrs 14 min Creating 50 Solr indexes on 1M news articles: 55 min Map: 7 min Reduce: 54 min 40

41 Open-sourced products and others in the open source pipeline iscore Content-based t personalization [pre-hadoop] Reconcile Coreference resolution software built on open source tools [with Cornell and U. Utah] Additional adaptation to Hadoop Dunk An elegant java annotation system that allows you to have the fields of a java object serialized (deserialized) to (from) an HBase table Simplifies queries, object construction, and map/reduce formulation 41

42 Questions? Disclaimer This document was prepared as an account of work sponsored by an agency of the United States government. Neither the United States government nor Lawrence Livermore National Security, LLC, nor any of their employees makes any warranty, expressed or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States government or Lawrence Livermore National Security, LLC. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States government or Lawrence Livermore National Security, LLC, and shall not be used for advertising or product endorsement purposes. 42