USING HDFS ON DISCOVERY CLUSTER TWO EXAMPLES - test1 and test2

Transcription

1 USING HDFS ON DISCOVERY CLUSTER TWO EXAMPLES - test1 and test2 (Using HDFS on Discovery Cluster for Discovery Cluster Users [email protected] if you have questions or need more clarifications. Nilay K. Roy, Ph.D) To use HDFS on Discovery Cluster login to any of the login nodes as usual. Make sure that your.bashrc has the following modules loaded. module whatis hadoop will give you the prerequisites. After login type module list to make sure the proper modules are loaded as shown below: Next using the hadoop-10g queue get an interactive-node. Every user is restricted to no more than 10 cores. The interactive nodes are the hadoop file system data nodes. Generally you do not need X-windows forwarding so bsub -Is -q hadoop-10g -n 1 /bin/bash will suffice. If you do need X-windows forwarding add the -XF tag for bsub -Is -XF -q hadoop-10g -n 1 /bin/bash. This is shown below. -n is number of cores requested. If you have a bad ssh connection and want a persistent one use screen available on the two login nodes discovery2 and discovery4. This way you can logout of the cluster while your job runs and login later without killing anything. Remember to detach from the screen session before logging out of the login nodes. You can reattach to the relevant screen after login. Further details are out-of-scope here. Please note that there is no time restriction on this queue. So when you are done exit from the login to the compute node and the LSF interactive session will automatically be killed. Check it again by using bjobs -w and if still running use bkill <job_ib>. Similarly exit all screen sessions if you are not doing anything in them.

2 Using HDFS on Discovery Cluster consists of three steps: 1) Compile hadoop code using Java and get the *.jar file. 2) Create your data directory on the HDFS file system following the Discovery Cluster best practice procedure as in /scratch. 3) Move (stage) the input data to the HDFS file system in your top level data folder. 4) Run your code. 5) Move your output and other data back to your /home/<my_neu_id> or /scratch/<my_neu_id> folder. 6) Exit your interactive session. Check your interactive job is killed on exit using bjobs w. 7) Exit screen if using it. 1. Compile Hadoop code: Example test1 Download the example file here - Now upload to your /home/<my_neu_id> or /scratch/<my_neu_id> and extract then compile using javac as shown below. Make a directory for the classes. In this example it is called wordcount_classes. Before compiling run export CLASSPATH=$(hadoop classpath):$classpath where hadoop will be in your path if you have loaded the modules correctly as shown above. You can ignore the deprecated API note. In the example java source file you will see import org.apache.hadoop.filecache.distributedcache;. This is for fully distributed Hadoop implementation with multi-level replication and fully distributed cache as in this case. In this case the replication and distributed cache is across three data nodes compute-2-004, compute-2-005, compute [nilay.roy@compute ~]$ tar -zxvf NKR_hadoop_example.tar.gz hadoop_test/ hadoop_test/test1/ hadoop_test/test1/wordcount.java hadoop_test/test1/file02 hadoop_test/test1/file01 [nilay.roy@compute ~]$ cd hadoop_test/ [nilay.roy@compute hadoop_test]$ cd test1 ls -la total 83 drwxr-xr-x 2 nilay.roy GID_nilay.roy 80 Aug 18 15:24. drwxr-xr-x 3 nilay.roy GID_nilay.roy 23 Aug 15 11:50.. -rw-r--r-- 1 nilay.roy GID_nilay.roy 24 Aug 15 19:32 file01 -rw-r--r-- 1 nilay.roy GID_nilay.roy 33 Aug 15 19:32 file02 -rw-r--r-- 1 nilay.roy GID_nilay.roy 4544 Aug 15 12:12 WordCount.java mkdir wordcount_classes export CLASSPATH=$(hadoop classpath):$classpath javac -d wordcount_classes WordCount.java Note: WordCount.java uses or overrides a deprecated API. Note: Recompile with -Xlint:deprecation for details. cd wordcount_classes/org/myorg [nilay.roy@compute myorg]$ ls -la total 109 drwxr-xr-x 2 nilay.roy GID_nilay.roy 156 Aug 18 15:44. drwxr-xr-x 3 nilay.roy GID_nilay.roy 23 Aug 18 15:44.. -rw-r--r-- 1 nilay.roy GID_nilay.roy 2671 Aug 18 15:44 WordCount.class -rw-r--r-- 1 nilay.roy GID_nilay.roy 4661 Aug 18 15:44 WordCount$Map.class -rw-r--r-- 1 nilay.roy GID_nilay.roy 983 Aug 18 15:44 WordCount$Map$Counters.class -rw-r--r-- 1 nilay.roy GID_nilay.roy 1611 Aug 18 15:44 WordCount$Reduce.class [nilay.roy@compute myorg]$ 2. Create your HDFS top-level directory: The rule followed for the top-level HDFS directory is <myneu_id>. This your id that you use to login into the cluster. You create the HDFS directory as follows: The command to use is hadoop fs -mkdir hdfs://discovery3:9000/tmp/nilay.roy where my my_neu_id is nilay.roy. Ignore the native-hadoop stack guard warnings. These will be fixed in later versions. We use a global Hadoop install via modules not a local per node install. Hence the WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where warnings.

3 hadoop fs -mkdir hdfs://discovery3:9000/tmp/nilay.roy 14/08/18 17:04:36 WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where You can now list the directory and also view it in the GUI. For GUI go to and click the Utilities tab on top. From the drop down select Browse the file system. From the command line - see below: hdfs dfs -ls hdfs://discovery3:9000/tmp/ 14/08/18 17:13:51 WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where Found 2 items drwxrwx--- - hadoopuser supergroup :34 hdfs://discovery3:9000/tmp/hadoop-yarn drwxr-xr-x - nilay.roy supergroup :04 hdfs://discovery3:9000/tmp/nilay.roy 3. Move (stage) the input data to the HDFS file system in your top level data folder: Now we can move the input file into the input directory. Create a local input directory for this file and put the files there and then copy the whole hadoop_test folder over. This is shown below: [nilay.roy@compute ~]$ hdfs dfs -put hadoop_test/ hdfs://discovery3:9000/tmp/nilay.roy/. 14/08/18 17:20:18 WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where [nilay.roy@compute ~]$ hdfs dfs -lsr hdfs://discovery3:9000/tmp/nilay.roy lsr: DEPRECATED: Please use 'ls -R' instead. 14/08/18 17:20:44 WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where drwxr-xr-x - nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test drwxr-xr-x - nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1 -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount.java -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/file01 -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/file02 drwxr-xr-x - nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/input -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/input/file01 -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/input/file02 drwxr-xr-x - nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes drwxr-xr-x - nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes/org drwxr-xr-x - nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes/org/myorg -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes/org/myorg/wordcount$map$counters.class -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes/org/myorg/wordcount$map.class -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes/org/myorg/wordcount$reduce.class -rw-r--r-- 3 nilay.roy supergroup :20 hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/wordcount_classes/org/myorg/wordcount.class [nilay.roy@compute ~]$ 4. Run your code: Now create the jar file to run as shown below using jar -cvf wordcount.jar -C wordcount_classes/.

4 jar -cvf wordcount.jar -C wordcount_classes/. added manifest adding: org/(in = 0) (out= 0)(stored 0%) adding: org/myorg/(in = 0) (out= 0)(stored 0%) adding: org/myorg/wordcount$map.class(in = 4661) (out= 2217)(deflated 52%) adding: org/myorg/wordcount.class(in = 2671) (out= 1289)(deflated 51%) adding: org/myorg/wordcount$map$counters.class(in = 983) (out= 504)(deflated 48%) adding: org/myorg/wordcount$reduce.class(in = 1611) (out= 648)(deflated 59%) ls -la total 114 drwxr-xr-x 4 nilay.roy GID_nilay.roy 169 Aug 18 17:27. drwxr-xr-x 3 nilay.roy GID_nilay.roy 23 Aug 15 11:50.. -rw-r--r-- 1 nilay.roy GID_nilay.roy 24 Aug 15 19:32 file01 -rw-r--r-- 1 nilay.roy GID_nilay.roy 33 Aug 15 19:32 file02 drwxr-xr-x 2 nilay.roy GID_nilay.roy 48 Aug 18 17:18 input drwxr-xr-x 3 nilay.roy GID_nilay.roy 21 Aug 18 15:44 wordcount_classes -rw-r--r-- 1 nilay.roy GID_nilay.roy 5799 Aug 18 17:27 wordcount.jar -rw-r--r-- 1 nilay.roy GID_nilay.roy 4544 Aug 15 12:12 WordCount.java Then run in on the HDFS file system as show below using: hadoop jar /home/nilay.roy/hadoop_test/test1/wordcount.jar org.myorg.wordcount hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/input hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/output Although the above seems intimidating what we are doing is simply giving the full path to the *.jar to run and the full paths for input and output in HDFS with the org.myorg.wordcount compiled class libraries. When you run this you will see output as shown below: hadoop jar /home/nilay.roy/hadoop_test/test1/wordcount.jar org.myorg.wordcount hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/input hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/output 14/08/18 17:35:44 WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where 14/08/18 17:35:44 INFO Configuration.deprecation: session.id is deprecated. Instead, use dfs.metrics.session-id 14/08/18 17:35:44 INFO jvm.jvmmetrics: Initializing JVM Metrics with processname=jobtracker, sessionid= 14/08/18 17:35:44 INFO jvm.jvmmetrics: Cannot initialize JVM Metrics with processname=jobtracker, sessionid= - already initialized 14/08/18 17:35:45 INFO mapred.fileinputformat: Total input paths to process : 2 14/08/18 17:35:45 INFO mapreduce.jobsubmitter: number of splits:2 14/08/18 17:35:46 INFO mapreduce.jobsubmitter: Submitting tokens for job: job_local _0001 ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::LOT OF OUTPUT OMITTED FOR CONCISENESS:::::::::::::::::::::::::::::::::::::::::::::::::::::::::: FILE: Number of write operations=0 HDFS: Number of bytes read=147 HDFS: Number of bytes written=67 HDFS: Number of read operations=25 Reduce output records=8 File Input Format Counters Bytes Read=57 File Output Format Counters Bytes Written=67 org.myorg.wordcount$map$counters INPUT_WORDS=9 5. Move data out from HDFS to your home directory: Browse your output data using the command shown below and then move it over using the hdfs dfs -get command. This cat example is shown below and the get is left as an exercise. You can then remove your data with the hdfs dfs - rmr command. This is also left as an exercise for the user.

5 hdfs dfs -cat hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test1/output/part /08/18 17:54:42 WARN util.nativecodeloader: Unable to load native-hadoop library for your platform... using builtin-java classes where Bye 1 Goodbye 1 Hadoop, 1 Hello 2 World! 1 World, 1 hadoop. 1 to 1 6. Exit your interactive session: Exit your interactive session. Check your interactive job is killed on exit using bjobs -w. The entire online Hadoop User Guide is here: File System Shell GUIDE is here: If you are not sure what you are doing, have questions, encounter problems, and require help with Hadoop, JAVA, PYTHON or any other issues contact ITS-Research Computing at [email protected]. Alternatively me at [email protected]. Please be mindful that this 50TB Hadoop cluster is a shared resource and must not be used for data storage. Once you are done with your work please clean up everything under your top - level HDFS directory that must be: hdfs://discovery.neu.edu:9000/tmp/<my_neu_id> - where <my_neu_id> is your Discovery Cluster Login ID and the same ID that you use to login to myneu DO NOT USE any other paths here.

6 Another Example test2 Download the example file here - Here is a more advanced example of secondary sort that shows of the power of Hadoop Map-Reduce. Here we sort the values coming into the Reducer of a Hadoop Map/Reduce (MR) Job. We extend the Java API for our use and do the following: 1.Use a composite key. 2.Extend org.apache.hadoop.mapreduce.partitioner. 3.Extend org.apache.hadoop.io.writablecomparator. The main part of the code that one should focus on is the use of the M/R API org.apache.hadoop.mapreduce.*. The problem Imagine we have stock data that looks like the following. Each line represents the value of a stock at a particular time. Each value in a line is delimited by a comma. The first value is the stock symbol (i.e. GOOG), the second value is the timestamp (i.e. the number of milliseconds since January 1, 1970, 00:00:00 GMT), and the third value is the stock s price. The data below is a toy data set. As you can see, there are 3 stock symbols: a, b, and c. The timestamps are also simple: 1, 2, 3, 4. The values are fake as well: 1.0, 2.0, 3.0, and 4.0. a, 1, 1.0 b, 1, 1.0 c, 1, 1.0 a, 2, 2.0 b, 2, 2.0 c, 2, 2.0 a, 3, 3.0 b, 3, 3.0 c, 3, 3.0 a, 4, 4.0 b, 4, 4.0 c, 4, 4.0 Let s say we want for each stock symbol (the reducer key input, or alternatively, the mapper key output), to order the values descendingly by timestamp when they come into the reducer. How do we sort the timestamp descendingly? This problem is known as secondary sorting. Hadoop s M/R platform sorts the keys, but not the values. (Note, Google s M/R platform explicitly supports secondary sorting, see Lin and Dyer 2010). A solution for secondary sorting Use a composite key A solution for secondary sorting involves doing multiple things. First, instead of simply emitting the stock symbol as the key from the mapper, we need to emit a composite key, a key that has multiple parts. The key will have the stock symbol and timestamp. The process for a M/R Job is as follows. (K1,V1) > Map > (K2,V2) (K2,List[V2]) > Reduce > (K3,V3) In the toy data above, K1 will be of type LongWritable, and V1 will be of type Text. Without secondary sorting, K2 will be of type Text and V2 will be of type DoubleWritable (we simply emit the stock symbol and price from the

7 mapper to the reducer). So, K2=symbol, and V2=price, or (K2,V2) = (symbol,price). However, if we emit such an intermediary key-value pair, secondary sorting is not possible. We have to emit a composite key, K2={symbol,timestamp. So the intermediary key-value pair is (K2,V2) = ({symbol,timestamp,price). Note that composite data structures, such as the composite key, is held within the curly braces. Our reducer simply outputs a K3 of type Text and V3 of type Text; (K3,V3) = (symbol, price). The complete M/R job with the new composite key is shown below. (LongWritable,Text) > Map > ({symbol,timestamp,price) ({symbol,timestamp,list[price]) > Reduce > (symbol,price) K2 is a composite key, but inside it, the symbol part/component is referred to as the natural key. It is the key which values will be grouped by. Use a composite key comparator The composite key comparator is where the secondary sorting takes place. It compares composite key by symbol ascendingly and timestamp descendingly. It is shown below. Notice here we sort based on symbol and timestamp. All the components of the composite key are considered. public class CompositeKeyComparator extends WritableComparator { protected CompositeKeyComparator() { public int compare(writablecomparable w1, WritableComparable w2) { StockKey k1 = (StockKey)w1; StockKey k2 = (StockKey)w2; int result = k1.getsymbol().compareto(k2.getsymbol()); if(0 == result) { result = -1* k1.gettimestamp().compareto(k2.gettimestamp()); return result; Use a natural key grouping comparator The natural key group comparator groups values together according to the natural key. Without this component, each K2={symbol,timestamp and its associated V2=price may go to different reducers. Notice here, we only consider the natural key. public class NaturalKeyGroupingComparator extends WritableComparator { protected NaturalKeyGroupingComparator() { public int compare(writablecomparable w1, WritableComparable w2) { StockKey k1 = (StockKey)w1; StockKey k2 = (StockKey)w2; return k1.getsymbol().compareto(k2.getsymbol());

8 Use a natural key partitioner The natural key partitioner uses the natural key to partition the data to the reducer(s). Again, note that here, we only consider the natural key. public class NaturalKeyPartitioner extends Partitioner<StockKey, DoubleWritable> public int getpartition(stockkey key, DoubleWritable val, int numpartitions) { int hash = key.getsymbol().hashcode(); int partition = hash % numpartitions; return partition; The M/R Job Once we define the Mapper, Reducer, natural key grouping comparator, natural key partitioner, composite key comparator, and composite key, in Hadoop s M/R API, we may configure the Job as follows. public class SsJob extends Configured implements Tool { public static void main(string[] args) throws Exception { ToolRunner.run(new Configuration(), new SsJob(), public int run(string[] args) throws Exception { Configuration conf = getconf(); Job job = new Job(conf, "secondary sort"); job.setjarbyclass(ssjob.class); job.setpartitionerclass(naturalkeypartitioner.class); job.setgroupingcomparatorclass(naturalkeygroupingcomparator.class); job.setsortcomparatorclass(compositekeycomparator.class); job.setmapoutputkeyclass(stockkey.class); job.setmapoutputvalueclass(doublewritable.class); job.setoutputkeyclass(text.class); job.setoutputvalueclass(text.class); job.setinputformatclass(textinputformat.class); job.setoutputformatclass(textoutputformat.class); job.setmapperclass(ssmapper.class); job.setreducerclass(ssreducer.class); job.waitforcompletion(true); return 0; Comments You need at least 4 new classes. The composite key class needs to hold the natural key and other data that you will sort on. The composite key comparator will perform the sorting of the keys (and thus values). The natural key grouping

9 comparator will group values based on the natural key. The natural key partitioner will send values with the same natural key to the same reducer. NOTES: After the job is executed the result is shown below: The job was run using the command: hadoop jar demo_classes.jar demo.ssjob -Dmapred.input.dir=hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test2/data - Dmapred.output.dir=hdfs://discovery3:9000/tmp/nilay.roy/hadoop_test/test2/result The jar file was created as shown below: The code was compiled using the command: javac -d demo_classes CompositeKeyComparator.java NaturalKeyGroupingComparator.java NaturalKeyPartitioner.java SsJob.java SsMapper.java SsReducer.java StockKey.java ============================================================================================ ============================================================================================ Nilay K Roy, Ph.D ========================================================== Nilay Roy, PhD Computational Physics, MS Computer Science Assistant Director - Research Computing, Information Technology Services Northeastern University, , 360 Huntington Avenue, Boston, MA [email protected] (C) (Preferred) / (O) Northeastern Research Computing Website: ==========================================================