On This page
On Related Pages
The selected root page could not be found.
In this example, we are not using MapReduce to its full potential. We are only using it to run jobs in parallel, one job for each chromosome. The phase algorithm from UW writes its output to local files instead of stdout. Many currently existing exe or apps can be run this way either as standard alone or pipeline.
Mapper
- Write the mapper script
~>cat phaseMapper.sh #!/bin/sh RESULT_BUCKET=s3://sagebio-YourUnixUsername/results # Send some bogus output to stdout so that mapreduce does not timeout # during phase processing since the phase algorithm does send output # to stdout on a regular basis perl -e 'while(! -e "./timetostop") { print "keepalive\n"; print STDERR "reporter:status:keepalive\n"; sleep 300; }' & while read S3_INPUT_FILE; do echo input to process ${S3_INPUT_FILE} 1>&2 # For debugging purposes, print out the files cached for us ls -la 1>&2 # Parse the s3 file path to get the file name LOCAL_INPUT_FILE=$(echo ${S3_INPUT_FILE} | perl -pe 'if (/^((s3[n]?):\/)?\/?([^:\/\s]+)((\/\w+)*\/)([\w\-\.]+[^#?\s]+)(.*)?(#[\w\-]+)?$/) {print "$6\n"};' | head -1) # Download the file from S3 echo hadoop fs -get ${S3_INPUT_FILE} ${LOCAL_INPUT_FILE} 1>&2 hadoop fs -get ${S3_INPUT_FILE} ${LOCAL_INPUT_FILE} 1>&2 # Run phase processing ./phase ${LOCAL_INPUT_FILE} ${LOCAL_INPUT_FILE}_out 100 1 100 # Upload the output files ls -la ${LOCAL_INPUT_FILE}*_out* 1>&2 for f in ${LOCAL_INPUT_FILE}*_out* do echo hadoop fs -put $f ${RESULT_BUCKET}/$LOCAL_INPUT_FILE/$f 1>&2 hadoop fs -put $f ${RESULT_BUCKET}/$LOCAL_INPUT_FILE/$f 1>&2 done echo processed ${S3_INPUT_FILE} 1>&2 echo 1>&2 echo 1>&2 done # Tell our background keepalive task to exit touch ./timetostop exit 0 - Upload the mapper script to S3 via the AWS console or s3curl
/work/platform/bin/s3curl.pl --id $USER --put phaseMapper.sh https://s3.amazonaws.com/sagetest-$USER/scripts/phaseMapper.sh
- Upload the phase binary to S3 too
/work/platform/bin/s3curl.pl --id $USER --put PHASE https://s3.amazonaws.com/sagetest-$USER/scripts/phase
Reducer
We do not need a reducer for this task. It is merely the output of the phase algorithm that we want. Therefore in the job configuration be sure to set "-jobconf", "mapred.reduce.tasks=0"
Input
- Write your input file
~>cat phaseInput.txt s3://sagebio-YourUnixUsername/input/ProSM_chrom_MT.phase.inp ... many more files, one per chromosome
- Upload your input file to S3 via the AWS console or s3curl
/work/platform/bin/s3curl.pl --id $USER --put phaseInput.txt https://s3.amazonaws.com/sagetest-$USER/input/phaseInput.txt
- Also upload all the data files referenced in phaseInput.txt to the location specified in that file.
Run the MapReduce Job
Job Configuration
- Write your job configuration. Note that you need to change the output location each time you run this!
~>cat phase.json [ { "Name": "MapReduce Step 1: Run Phase", "ActionOnFailure": "CANCEL_AND_WAIT", "HadoopJarStep": { "Jar": "/home/hadoop/contrib/streaming/hadoop-streaming.jar", "Args": [ "-input", "s3n://sagebio-YourUnixUsername/input/phaseInput.txt", "-output", "s3n://sagebio-YourUnixUsername/output/phaseTry1", "-mapper", "s3n://sagebio-YourUnixUsername/scripts/phaseMapper.sh", "-cacheFile", "s3n://sagebio-YourUnixUsername/scripts/phase#phase", "-jobconf", "mapred.reduce.tasks=0", "-jobconf", "mapred.task.timeout=604800000", ] } } ] - Put it on one of the shared servers sodo/ballard/belltown.
If you find that your mapper tasks are not getting balanced evenly across your fleet, you can add lines like the following to your job config:
"-jobconf", "mapred.map.tasks=26", "-jobconf", "mapred.tasktracker.map.tasks.maximum=2",
Start the MapReduce cluster
- ssh to one of the shared servers sodo/ballard/belltown
- Kick of the Elastic Map Reduce Job. This will start 14 hosts: one for the master and 13 for the slaves running the map tasks.
~>elastic-mapreduce --credentials ~/.ssh/$USER-credentials.json --create \ --enable-debugging --bootstrap-action s3://elasticmapreduce/bootstrap-actions/configurations/latest/memory-intensive \ --master-instance-type=m1.small --slave-instance-type=c1.medium --num-instances=14 --json phase.json --name phaseTry1 Created job flow j-GA47B7VD991Q
Check on the job status
If something is misconfigured, it will fail in a minute or two. Check on the job status and make sure it is running.
~>elastic-mapreduce --credentials ~/.ssh/$USER-credentials.json --list --jobflow j-GA47B7VD991Q j-GA47B7VD991Q RUNNING ec2-174-129-134-200.compute-1.amazonaws.com filesysTry1 RUNNING MapReduce Step 1: Run Phase
If there were any errors, make corrections and resubmit the job step
~>elastic-mapreduce --credentials ~/.ssh/$USER-credentials.json --json phase.json --jobflow j-GA47B7VD991Q Added jobflow steps
Get your results
Look in your S3 bucket for the results.
How to gain more parallelization by splitting your input files into multiple chunks
You can find the Python 2.7 script to split the scripts in subversion: phaseSplit.py
- Usage:
~/>python2.7 phaseSplit.py --help usage: phaseSplit.py [-h] --phaseInputFile PHASEINPUTFILE [--minColumnsPerFile MINCOLUMNSPERFILE] [--columnOverlap COLUMNOVERLAP] Split phase input files into smaller chunks optional arguments: -h, --help show this help message and exit --phaseInputFile PHASEINPUTFILE, -p PHASEINPUTFILE the file path to the phase input file to be split --minColumnsPerFile MINCOLUMNSPERFILE, -m MINCOLUMNSPERFILE the minimum number of columns to output per file --columnOverlap COLUMNOVERLAP, -o COLUMNOVERLAP the number of columns to overlap with each file
- How to run it:
~/>python2.7 phaseSplit.py -p ProSM_chrom_21.phase.inp Sample 0 chunk 0 startColumn 0 endColumn 100 Sample 0 chunk 1 startColumn 80 endColumn 180 Sample 0 chunk 2 startColumn 160 endColumn 260 Sample 0 chunk 3 startColumn 240 endColumn 340 ... Sample 73 chunk 155 startColumn 12400 endColumn 12500 Sample 73 chunk 156 startColumn 12480 endColumn 12564 SUCCESS: ProSM_chrom_21.phase.inp and ProSM_chrom_21.phase.inp_sanityCheck are equivalent