Why use the pipeline to run Monte-Carlo jobs?
The pipeline offers a number of advantages when running large numbers of Monte-Carlo jobs:
- Provides an easy way to submit a large number of jobs
- Maintains record of all jobs run, including links to log files and other files produced during the job
- Makes it easy to rerun any jobs that fail due to quirks in the SLAC batch system.
- Makes it easy to register output datasets in the data catalog, which in turn makes it easy to keep track of what MC data is available.
- Provides a web interface to allow the status of jobs to be monitored from anywhere.
Example Monte-Carlo task
Normally to run a set of Monte-Carlo jobs it is necessary to define a pipeline "Task". A task consists of an arbitrary graph of batch jobs and "scriptlets" to be run, however a typical Monte-Carlo task consists of just two steps:
In this example the two steps are:
- runMonteCarlo - a batch job that is run to execute the simulation program and generate the output file
- register-ds - a scriptlet which is run if the batch job finishes successfully that registers the output file in the data catalog.
Typically when generating large MC event samples it is necessary to run many MC jobs each of which generates some number of events. In the pipeline this is achieved by creating many streams within a given task. In the case of this example Monte-Carlo task each stream will run one batch job followed by one registration scriplet.
Running jobs
To actually run some jobs it is necessary to call the createStream function of the pipeline. The easiest way to do this is to use the pipeline createStream command when logged in to SLAC unix (e.g. noric).
~exodata/pipeline/prod/pipeline createStream [-options] <taskname> [file1 [file2 [...]]...]
where possible options include:
--stream <Stream ID=-1> |
Integer stream identifier. Auto assigned if option not specified. |
||
--nStreams <Number of Streams=1> |
Number of streams to create, not valid if Stream ID specified |
||
<ac:structured-macro ac:name="unmigrated-wiki-markup" ac:schema-version="1" ac:macro-id="c5282bc0-c21c-4601-8f45-107fba96b6a7"><ac:plain-text-body><![CDATA[ |
--define <name=value> |
Define a variable. Syntax is "name=value[,name2=value2,...]" |
]]></ac:plain-text-body></ac:structured-macro> |
For example to create 10 streams of the EXOMCBackground task defined above, overriding the default value of the MAXEVENTS variable we would use the following command:
~exodata/pipeline/prod/pipeline createStream --define MAXEVENTS=100000 --nStreams 10 EXOMCBackground
Monitoring jobs using the pipeline web interface
The pipeline web interface can be access from the EXO data portal at:
http://exo-data.slac.stanford.edu/
The web interface allows monitoring of the status of tasks and streams, viewing the log files of running or completed jobs and "rolling back" (rerunning) any failed jobs. It is also possible to view plots of how much CPU time jobs took, how much many jobs were running at a given time etc.
Also linked from the EXO data portal is the data catalog web interface. This allows the output files registered by the MC jobs to be viewed and downloaded.
Defining a task
To create a pipeline task it is necessary to write an XML configuration file. The key elements of the XML configuration file for the task above (with some details initially left out) are shown here:
<?xml version="1.0" encoding="UTF-8"?>
<pipeline xmlns="http://glast-ground.slac.stanford.edu/pipeline"
xmlns:xs="http://www.w3.org/2001/XMLSchema-instance"
xs:schemaLocation="http://glast-ground.slac.stanford.edu/pipeline
http://srs.slac.stanford.edu/Pipeline-II/schemas/2.0/pipeline.xsd">
<task name="EXOMCBackground" type="EXO" version="1.11">
<notation>A generic task for running EXO MC backgrounds</notation>
<variables>
<var name="EXODIR">/nfs/slac/g/exo</var>
<var name="EXOBASE">${EXODIR}/software/builds/trunk</var>
<var name="BATCHOPTIONS">-R "select[-rhel30] rusage[scratch=1]"</var>
<var name="CORE_LIMIT">1024</var>
<var name="MAXEVENTS">10000</var>
<var name="PRINTMODULO">${MAXEVENTS/100}</var>
<var name="INITIALSEED">pipeline.stream%100000</var>
<var name="MAXCPU">${MAXEVENTS/10}</var>
<var name="MAXMEM">1000</var>
<var name="SOURCE_VOLUME">HFE</var>
<var name="SOURCE_ION">k</var>
<var name="OUTPUT_DIR">/nfs/slac/g/exo/exo_data/data/MC/backgrounds/TestBkgdMC/${SOURCE_ION}/${SOURCE_VOLUME}</var>
<var name="OUTPUT_FORMAT">MC-background-%06d.root</var>
<var name="OUTPUT_NAME">${format(pipeline.stream,OUTPUT_FORMAT)}</var>
<var name="OUTPUT_FILE">${OUTPUT_DIR}/${OUTPUT_NAME}</var>
<var name="DATACAT_DIR">EXO/TestBkgdMC/${SOURCE_ION}</var>
<var name="DATACAT_GROUP">${SOURCE_VOLUME}</var>
</variables>
<process name="runMonteCarlo">
<job batchOptions="${BATCHOPTIONS}" maxCPU="${MAXCPU}" maxMemory="${MAXMEM}">
...
</job>
</process>
<process name="register-ds">
<notation>Register datasets created in this task</notation>
<script>
...
</script>
<depends>
<after process="runMonteCarlo"/>
</depends>
</process>
</task>
</pipeline>
In this file the <task> element defines the name and version # of the task, as well as the task type. <notation> just indicates a comment describing the task. The <variables> section defines a set of variables which will be used elsewhere in the task. The values given to the variables are defaults which can be overridden for any specific stream when the stream is created. Note that variables can be defined in terms of other variables by using ${expression} syntax.
The two steps of the task are each defined using a <process> element. The first <process> contains a <job> element indicating it is a batch job (the body of the job is not omitted for the moment). Note again how variables defined earlier in the task are used to provide values for some parameters using the ${expression} syntax. The second <process> contains a <script> element indicating that it is a scriptlet (again the body of the scriptlet is omitted for the time being). The <depends> element indicates that the scriptlet should only run after the batch job successfully completes.
Now lets look at the parts which we initially missed out. First the body of the batch job, which by default is written as a bash script:
ulimit -c ${CORE_LIMIT} # Limit core dumps
set -e # exit on error
# Create a scratch area to write the output to
export SCRATCH_DIR=/scratch/exo/${PIPELINE_PROCESSINSTANCE}
mkdir -p ${SCRATCH_DIR}
gotEXIT()
{
rm -rf ${SCRATCH_DIR}
}
trap gotEXIT EXIT
source ${EXOBASE}/setup.sh
# Create background.exe
cat > background.exo <<EOF
use exosim rec toutput
/exosim/macro background.mac
/exosim/filter true
printmodulo ${PRINTMODULO}
/exosim/initial_seed ${INITIALSEED}
/exosim/run_number ${PIPELINE_STREAM}
maxevents ${MAXEVENTS}
/toutput/file ${SCRATCH_DIR}/output.root
begin
exit
EOF
case ${SOURCE_ION} in
k)
GPS_ION="19 40 0 0"
;;
th)
GPS_ION="90 232 0 0"
;;
u)
GPS_ION="92 238 0 0"
;;
*)
echo "Unknown ION ${SOURCE_ION}"
exit 1
esac
case ${SOURCE_VOLUME} in
HFE)
HALFZ=72.5
RADIUS=75.0
;;
InnnerCryo)
HALFZ=74.5
RADIUS=78.0
;;
*)
echo "Unknown volume ${SOURCE_VOLUME}"
exit 1
esac
# Create background.mac
cat > background.mac <<EOF
/digitizer/wireNoise 800.000000
/digitizer/APDNoise 2000.000000
/digitizer/LXeEnergyRes 0.015000
/event/LXeEventsOnly true
/event/digitizeWires true
/event/digitizeAPDs true
/gps/pos/type Volume
/gps/pos/shape Cylinder
/gps/pos/halfz ${HALFZ} cm
/gps/pos/radius ${RADIUS} cm
/gps/pos/centre 0.0 0.0 0.0 cm
/gps/pos/confine ${SOURCE_VOLUME}
/gps/energy 0 keV
/gps/particle ion
/gps/ion ${GPS_ION}
/grdm/analogueMC 1
EOF
EXOAnalysis background.exo
mkdir -p ${OUTPUT_DIR}
cp -pv ${SCRATCH_DIR}/output.root ${OUTPUT_FILE}
There are a few points worth noting
- All of the variables defined earlier in the task are passed to the batch job as environment variables and ca be referred to using the bash ${VARIABLE} syntax.
- The bash file creates a scratch folder at the top of the job, and registers a cleanup trap to delete the scratch area at the end of the job. The output data is written to the scratch area and copied to its final location if the job completes successfully. This is recommended practice since if many batch jobs write simultaneously to the same NFS file server it will likely become overloaded and fail horribly.
- The .mac file and .exo file required for running EXOAnalysis are generated on the fly, substituting in settings from the variables defined earlier. Note also the use of some special pipeline variables:
- ${PIPELINE_STREAM} -- an id assigned when each stream is created. The id will be unique within this task and normally starts at 0 and increments for each subsequent stream
- ${PIPELINE_PROCESSINSTANCE} -- similar to pipeline stream this is a unique id associated with a stream, but this id is unique across all tasks, and is normally a large ugly number
Finally lets look at the body of the scriptlet which is used to register the output dataset. The scriptlet is written in python:
metaData = {'nGeneratedEvents':MAXEVENTS,'SourceVolume':SOURCE_VOLUME,'SourceIon':SOURCE_ION}
dsNew = datacatalog.newDataset(OUTPUT_NAME, "root", "EXOROOT", DATACAT_DIR, DATACAT_GROUP, "SLAC", OUTPUT_FILE)
datacatalog.registerDataset(dsNew, metaData)
Again a few things worth noting:
- The datacatalog allows arbitrary meta-data to be associated with datasets. In this case the meta-data is defined as a python dictionary
- The newDataset method has many arguments which need some explanation.
Putting everything together the full XML file for the task is:
<?xml version="1.0" encoding="UTF-8"?>
<pipeline xmlns="http://glast-ground.slac.stanford.edu/pipeline"
xmlns:xs="http://www.w3.org/2001/XMLSchema-instance"
xs:schemaLocation="http://glast-ground.slac.stanford.edu/pipeline
http://srs.slac.stanford.edu/Pipeline-II/schemas/2.0/pipeline.xsd">
<task name="EXOMCBackground" type="EXO" version="1.11">
<notation>A generic task for running EXO MC backgrounds</notation>
<variables>
<var name="EXODIR">/nfs/slac/g/exo</var>
<var name="EXOBASE">${EXODIR}/software/builds/trunk</var>
<var name="BATCHOPTIONS">-R "select[-rhel30] rusage[scratch=1]"</var>
<var name="CORE_LIMIT">1024</var>
<var name="MAXEVENTS">10000</var>
<var name="PRINTMODULO">${MAXEVENTS/100}</var>
<var name="INITIALSEED">pipeline.stream%100000</var>
<var name="MAXCPU">${MAXEVENTS/10}</var>
<var name="MAXMEM">1000</var>
<var name="SOURCE_VOLUME">HFE</var>
<var name="SOURCE_ION">k</var>
<var name="OUTPUT_DIR">/nfs/slac/g/exo/exo_data/data/MC/backgrounds/TestBkgdMC/${SOURCE_ION}/${SOURCE_VOLUME}</var>
<var name="OUTPUT_FORMAT">MC-background-%06d.root</var>
<var name="OUTPUT_NAME">${format(pipeline.stream,OUTPUT_FORMAT)}</var>
<var name="OUTPUT_FILE">${OUTPUT_DIR}/${OUTPUT_NAME}</var>
<var name="DATACAT_DIR">EXO/TestBkgdMC/${SOURCE_ION}</var>
<var name="DATACAT_GROUP">${SOURCE_VOLUME}</var>
</variables>
<process name="runMonteCarlo">
<job batchOptions="${BATCHOPTIONS}" maxCPU="${MAXCPU}" maxMemory="${MAXMEM}"><![CDATA[
ulimit -c ${CORE_LIMIT} # Limit core dumps
set -e # exit on error
# Create a scratch area to write the output to
export SCRATCH_DIR=/scratch/exo/${PIPELINE_PROCESSINSTANCE}
mkdir -p ${SCRATCH_DIR}
gotEXIT()
{
rm -rf ${SCRATCH_DIR}
}
trap gotEXIT EXIT
source ${EXOBASE}/setup.sh
# Create background.exe
cat > background.exo <<EOF
use exosim rec toutput
/exosim/macro background.mac
/exosim/filter true
printmodulo ${PRINTMODULO}
/exosim/initial_seed ${INITIALSEED}
/exosim/run_number ${PIPELINE_STREAM}
maxevents ${MAXEVENTS}
/toutput/file ${SCRATCH_DIR}/output.root
begin
exit
EOF
case ${SOURCE_ION} in
k)
GPS_ION="19 40 0 0"
;;
th)
GPS_ION="90 232 0 0"
;;
u)
GPS_ION="92 238 0 0"
;;
*)
echo "Unknown ION ${SOURCE_ION}"
exit 1
esac
case ${SOURCE_VOLUME} in
HFE)
HALFZ=72.5
RADIUS=75.0
;;
InnnerCryo)
HALFZ=74.5
RADIUS=78.0
;;
*)
echo "Unknown volume ${SOURCE_VOLUME}"
exit 1
esac
# Create background.mac
cat > background.mac <<EOF
/digitizer/wireNoise 800.000000
/digitizer/APDNoise 2000.000000
/digitizer/LXeEnergyRes 0.015000
/event/LXeEventsOnly true
/event/digitizeWires true
/event/digitizeAPDs true
/gps/pos/type Volume
/gps/pos/shape Cylinder
/gps/pos/halfz ${HALFZ} cm
/gps/pos/radius ${RADIUS} cm
/gps/pos/centre 0.0 0.0 0.0 cm
/gps/pos/confine ${SOURCE_VOLUME}
/gps/energy 0 keV
/gps/particle ion
/gps/ion ${GPS_ION}
/grdm/analogueMC 1
EOF
EXOAnalysis background.exo
mkdir -p ${OUTPUT_DIR}
cp -pv ${SCRATCH_DIR}/output.root ${OUTPUT_FILE}
]]>
</job>
</process>
<process name="register-ds">
<notation>Register datasets created in this task</notation>
<script><![CDATA[
metaData = {'nGeneratedEvents':MAXEVENTS,'SourceVolume':SOURCE_VOLUME,'SourceIon':SOURCE_ION}
dsNew = datacatalog.newDataset(OUTPUT_NAME, "root", "EXOROOT", DATACAT_DIR, DATACAT_GROUP, "SLAC", OUTPUT_FILE)
datacatalog.registerDataset(dsNew, metaData)
]]>
</script>
<depends>
<after process="runMonteCarlo"/>
</depends>
</process>
</task>
</pipeline>
Once a new task has been defined it can be uploaded using either the pipeline web interface (on the Admin page) or from the pipeline command on SLAC unix:
~exodata/pipeline/prod/pipeline load <xml-file>
Note that each file to be uploaded must have a unique task name and version number, so when uploading new version of a task it is necessary to increment the version number in the <task> element.
List of available variables for EXOMCBackground (v1.12 and higher)
SOURCE_ION |
Description |
|---|---|
k |
Potassium 40 |
th |
Thorium-232 chain |
u |
Uranium-238 chain |
rn_220 |
Radon-220 chain |
rn_222 |
Radon-222 chain |
SOURCE_VOLUME |
Description |
|---|---|
ActiveLXe |
xenon volume |
LXeVessel |
xenon vessel |
HFE |
hfe |
InnerCryo |
inner cryostat |
OuterCryo |
outer cryostat |
LeadShield |
lead |