Name

pegasus-plan — runs Pegasus to generate the executable workflow

Synopsis

pegasus-plan [-v] [-q] [-V] [-h]
             [-Dprop=value…]] [-b prefix]
             [--conf propsfile]
             [-c cachefile[,cachefile…]] [--cleanup cleanup strategy ]
             [-C style[,style…]]
             [--dir dir]
             [--force] [--force-replan]
             [--inherited-rc-files file1[,file2…]] [-j prefix]
             [-n][-I input-dir1[,input-dir2…]][-O output-dir] [-o site]
             [-s site1[,site2…]]
             [--staging-site s1=ss1[,s2=ss2[..]]
             [--randomdir[=dirname]]
             [--relative-dir dir]
             [--relative-submit-dir dir]
             -d daxfile

Description

The pegasus-plan command takes in as input the DAX and generates an executable workflow usually in form of condor submit files, which can be submitted to an execution site for execution.

As part of generating an executable workflow, the planner needs to discover:

data

The Pegasus Workflow Planner ensures that all the data required for the execution of the executable workflow is transferred to the execution site by adding transfer nodes at appropriate points in the DAG. This is done by looking up an appropriate Replica Catalog to determine the locations of the input files for the various jobs. By default, a file based replica catalog is used.

The Pegasus Workflow Planner also tries to reduce the workflow, unless specified otherwise. This is done by deleting the jobs whose output files have been found in some location in the Replica Catalog. At present no cost metrics are used. However preference is given to a location corresponding to the execution site

The planner can also add nodes to transfer all the materialized files to an output site. The location on the output site is determined by looking up the site catalog file, the path to which is picked up from the pegasus.catalog.site.file property value.

executables
The planner looks up a Transformation Catalog to discover locations of the executables referred to in the executable workflow. Users can specify INSTALLED or STAGEABLE executables in the catalog. Stageable executables can be used by Pegasus to stage executables to resources where they are not pre-installed.
resources
The layout of the sites, where Pegasus can schedule jobs of a workflow are described in the Site Catalog. The planner looks up the site catalog to determine for a site what directories a job can be executed in, what servers to use for staging in and out data and what jobmanagers (if applicable) can be used for submitting jobs.

The data and executable locations can now be specified in DAX’es conforming to DAX schema version 3.2 or higher.

Options

Any option will be displayed with its long options synonym(s).

-Dproperty=value
The -D option allows an experienced user to override certain properties which influence the program execution, among them the default location of the user’s properties file and the PEGASUS home location. One may set several CLI properties by giving this option multiple times. The -D option(s) must be the first option on the command line. A CLI property take precedence over the properties file property of the same key.
-d file , --dax file
The DAX is the XML input file that describes an abstract workflow. This is a mandatory option, which has to be used.
-b prefix , --basename prefix
The basename prefix to be used while constructing per workflow files like the dagman file (.dag file) and other workflow specific files that are created by Condor. Usually this prefix, is taken from the name attribute specified in the root element of the dax files.
-c file[,file,…] , --cache file[,file,…]

A comma separated list of paths to replica cache files that override the results from the replica catalog for a particular LFN.

Each entry in the cache file describes a LFN , the corresponding PFN and the associated attributes. The site attribute should be specified for each entry.

LFN_1 PFN_1 site=[site handle 1]
LFN_2 PFN_2 site=[site handle 2]
 ...
LFN_N PFN_N [site handle N]

To treat the cache files as supplemental replica catalogs set the property pegasus.catalog.replica.cache.asrc to true. This results in the mapping in the cache files to be merged with the mappings in the replica catalog. Thus, for a particular LFN both the entries in the cache file and replica catalog are available for replica selection.

-C style[,style,…] , --cluster style[,style,…]

Comma-separated list of clustering styles to apply to the workflow. This mode of operation results in clustering of n compute jobs into a larger jobs to reduce remote scheduling overhead. You can specify a list of clustering techniques to recursively apply them to the workflow. For example, this allows you to cluster some jobs in the workflow using horizontal clustering and then use label based clustering on the intermediate workflow to do vertical clustering.

The clustered jobs can be run at the remote site, either sequentially or by using MPI. This can be specified by setting the property pegasus.job.aggregator. The property can be overridden by associating the PEGASUS profile key collapser either with the transformation in the transformation catalog or the execution site in the site catalog. The value specified (to the property or the profile), is the logical name of the transformation that is to be used for clustering jobs. Note that clustering will only happen if the corresponding transformations are catalogued in the transformation catalog.

PEGASUS ships with a clustering executable pegasus-cluster that can be found in the $PEGASUS_HOME/bin directory. It runs the jobs in the clustered job sequentially on the same node at the remote site.

In addition, an MPI based clustering tool called pegasus-mpi-cluster', is also distributed and can be found in the bin directory. pegasus-mpi-cluster can also be used in the sharedfs setup and needs to be compiled against the remote site MPI install. directory. The wrapper is run on every MPI node, with the first one being the master and the rest of the ones as workers.

By default, pegasus-cluster is used for clustering jobs unless overridden in the properties or by the pegasus profile key collapser.

The following type of clustering styles are currently supported:

  • horizontal is the style of clustering in which jobs on the same level are aggregated into larger jobs. A level of the workflow is defined as the greatest distance of a node, from the root of the workflow. Clustering occurs only on jobs of the same type i.e they refer to the same logical transformation in the transformation catalog.

    Horizontal Clustering can operate in one of two modes. a. Job count based.

    The granularity of clustering can be specified by associating either the PEGASUS profile key clusters.size or the PEGASUS profile key clusters.num with the transformation.

    The clusters.size key indicates how many jobs need to be clustered into the larger clustered job. The clusters.num key indicates how many clustered jobs are to be created for a particular level at a particular execution site. If both keys are specified for a particular transformation, then the clusters.num key value is used to determine the clustering granularity.

    1. Runtime based.

      To cluster jobs according to runtimes user needs to set one property and two profile keys. The property pegasus.clusterer.preference must be set to the value runtime. In addition user needs to specify two Pegasus profiles. a. clusters.maxruntime which specifies the maximum duration for which the clustered job should run for. b. job.runtime which specifies the duration for which the job with which the profile key is associated, runs for. Ideally, clusters.maxruntime should be set in transformation catalog and job.runtime should be set for each job individually.

  • label is the style of clustering in which you can label the jobs in your workflow. The jobs with the same level are put in the same clustered job. This allows you to aggregate jobs across levels, or in a manner that is best suited to your application.

    To label the workflow, you need to associate PEGASUS profiles with the jobs in the DAX. The profile key to use for labeling the workflow can be set by the property pegasus.clusterer.label.key. It defaults to label, meaning if you have a PEGASUS profile key label with jobs, the jobs with the same value for the pegasus profile key label will go into the same clustered job.

--cleanup cleanup strategy

The cleanup strategy to be used for workflows. Pegasus can add cleanup jobs to the executable workflow that can remove files and directories during the workflow execution. The default strategy is inplace .

The following type of cleanup strategies are currently supported:

  • none disables cleanup altogether. The planner does not add any cleanup jobs in the executable workflow whatsoever.
  • leaf the planner adds a leaf cleanup node per staging site that removes the directory created by the create dir job in the workflow.
  • inplace the planner adds in addition to leaf cleanup nodes, cleanup nodes per level of the workflow that remove files no longer required during execution. For example, an added cleanup node will remove input files for a particular compute job after the job has finished successfully.
  • constraint the planner adds in addition to leaf cleanup nodes, cleanup nodes to constraint the amount of storage space used by a workflow. The added cleanup node guarantees limits on disk usage.

    By default, for hierarchal workflows the inplace cleanup is always turned off. This is because the cleanup algorithm ( InPlace ) does not work across the sub workflows. For example, if you have two DAX jobs in your top level workflow and the child DAX job refers to a file generated during the execution of the parent DAX job, the InPlace cleanup algorithm when applied to the parent dax job will result in the file being deleted, when the sub workflow corresponding to parent DAX job is executed. This would result in failure of sub workflow corresponding to the child DAX job, as the file deleted is required to present during it’s execution.

    In case there are no data dependencies across the dax jobs, then yes you can enable the InPlace algorithm for the sub dax’es . To do this you can set the property

    pegasus.file.cleanup.scope deferred

    This will result in cleanup option to be picked up from the arguments for the DAX job in the top level DAX.

--conf propfile
The path to properties file that contains the properties planner needs to use while planning the workflow. Defaults to pegasus.properties file in the current working directory, if no conf option is specified.
--dir dir

The base directory where you want the output of the Pegasus Workflow Planner usually condor submit files, to be generated. Pegasus creates a directory structure in this base directory on the basis of username, VO Group and the label of the workflow in the DAX.

By default the base directory is the directory from which one runs the pegasus-plan command.

-f , --force
This bypasses the reduction phase in which the abstract DAG is reduced, on the basis of the locations of the output files returned by the replica catalog. This is analogous to a make style generation of the executable workflow.
--force-replan
By default, for hierarichal workflows if a DAX job fails, then on job retry the rescue DAG of the associated workflow is submitted. This option causes Pegasus to replan the DAX job in case of failure instead.
-g , --group
The VO Group to which the user belongs to.
-h , --help
Displays all the options to the pegasus-plan command.
--inherited-rc-files file[,file,…]
A comma separated list of paths to replica files. Locations mentioned in these have a lower priority than the locations in the DAX file. This option is usually used internally for hierarchical workflows, where the file locations mentioned in the parent (encompassing) workflow DAX, passed to the sub workflows (corresponding) to the DAX jobs.
-I , --input-dir

A comma separated list of input directories on the submit host where the input files reside. This internally loads a Directory based Replica Catalog backend, that constructs does a directory listing to create the LFN→PFN mappings for the files in the input directory. You can specify additional properties either on the command line or the properties file to control the site attribute and url prefix associated with the mappings.

pegasus.catalog.replica.directory.site specifies the site attribute to associate with the mappings. Defaults to local

pegasus.catalog.replica.directory.url.prefix specifies the URL prefix to use while constructing the PFN. Defaults to file://

-j prefix , --job-prefix prefix
The job prefix to be applied for constructing the filenames for the job submit files.
-n , --nocleanup
This option is deprecated. Use --cleanup none instead.
-o site , --output-site site

The output site to which the output files of the DAX are transferred to.

By default the materialized data remains in the working directory on the execution site where it was created. Only those output files are transferred to an output site for which transfer attribute is set to true in the DAX.

-O output directory , --output-dir output directory

The output directory to which the output files of the DAX are transferred to.

If -o is specified the storage directory of the site specified as the output site is updated to be the directory passed. If no output site is specified, then this option internally sets the output site to local with the storage directory updated to the directory passed.

-q , --quiet
Decreases the logging level.
-r[dirname] , --randomdir[=dirname]

Pegasus Worfklow Planner adds create directory jobs to the executable workflow that create a directory in which all jobs for that workflow execute on a particular site. The directory created is in the working directory (specified in the site catalog with each site).

By default, Pegasus duplicates the relative directory structure on the submit host on the remote site. The user can specify this option without arguments to create a random timestamp based name for the execution directory that are created by the create dir jobs. The user can can specify the optional argument to this option to specify the basename of the directory that is to be created.

The create dir jobs refer to the dirmanager executable that is shipped as part of the PEGASUS worker package. The transformation catalog is searched for the transformation named pegasus::dirmanager for all the remote sites where the workflow has been scheduled. Pegasus can create a default path for the dirmanager executable, if PEGASUS_HOME environment variable is associated with the sites in the site catalog as an environment profile.

--relative-dir dir
The directory relative to the base directory where the executable workflow it to be generated and executed. This overrides the default directory structure that Pegasus creates based on username, VO Group and the DAX label.
--relative-submit-dir dir
The directory relative to the base directory where the executable workflow it to be generated. This overrides the default directory structure that Pegasus creates based on username, VO Group and the DAX label. By specifying --relative-dir and --relative-submit-dir you can have different relative execution directory on the remote site and different relative submit directory on the submit host.
-s site[,site,…] , --sites site[,site,…]

A comma separated list of execution sites on which the workflow is to be executed. Each of the sites should have an entry in the site catalog, that is being used.

In case this option is not specified, all the sites in the site catalog other than site local are picked up as candidates for running the workflow.

--staging-site s1=ss1[,s2=ss2[..]]

A comma separated list of key=value pairs , where the key is the execution site and value is the staging site for that execution site.

In case of running on a shared filesystem, the staging site is automatically associated by the planner to be the execution site. If only a value is specified, then that is taken to be the staging site for all the execution sites. e.g --staging-site local means that the planner will use the local site as the staging site for all jobs in the workflow.

-s , --submit
Submits the generated executable workflow using pegasus-run script in $PEGASUS_HOME/bin directory. By default, the Pegasus Workflow Planner only generates the Condor submit files and does not submit them.
-v , --verbose

Increases the verbosity of messages about what is going on. By default, all FATAL, ERROR, CONSOLE and WARN messages are logged. The logging hierarchy is as follows:

  1. FATAL
  2. ERROR
  3. CONSOLE
  4. WARN
  5. INFO
  6. CONFIG
  7. DEBUG
  8. TRACE

For example, to see the INFO, CONFIG and DEBUG messages additionally, set -vvv.

-V , --version
Displays the current version number of the Pegasus Workflow Management System.

Return Value

If the Pegasus Workflow Planner is able to generate an executable workflow successfully, the exitcode will be 0. All runtime errors result in an exitcode of 1. This is usually in the case when you have misconfigured your catalogs etc. In the case of an error occurring while loading a specific module implementation at run time, the exitcode will be 2. This is usually due to factory methods failing while loading a module. In case of any other error occurring during the running of the command, the exitcode will be 1. In most cases, the error message logged should give a clear indication as to where things went wrong.

Controlling pegasus-plan Memory Consumption

pegasus-plan will try to determine memory limits automatically using factors such as total system memory and potential memory limits (ulimits). The automatic limits can be overridden by setting the JAVA_HEAPMIN and JAVA_HEAPMAX environment variables before invoking pegasus-plan. The values are in megabytes. As a rule of thumb, JAVA_HEAPMIN can be set to half of the value of JAVA_HEAPMAX.

Pegasus Properties

This is not an exhaustive list of properties used. For the complete description and list of properties refer to $PEGASUS_HOME/doc/advanced-properties.pdf

pegasus.selector.site
Identifies what type of site selector you want to use. If not specified the default value of Random is used. Other supported modes are RoundRobin and NonJavaCallout that calls out to a external site selector.
pegasus.catalog.replica

Specifies the type of replica catalog to be used.

If not specified, then the value defaults to File.

pegasus.catalog.replica.url
Contact string to access the replica catalog. In case of File it is path to the file based replica catalog. If not specified, then default value of $PWD/rc.txt is used for the default File based Replica Catalog.
pegasus.dir.exec
A suffix to the workdir in the site catalog to determine the current working directory. If relative, the value will be appended to the working directory from the site.config file. If absolute it constitutes the working directory.
pegasus.catalog.transformation
Specifies the type of transformation catalog to be used. One can use only a file based transformation catalog, with the value as Text.
pegasus.catalog.transformation.file

The location of file to use as transformation catalog.

If not specified, then the default location of $PWD/tc.txt

pegasus.catalog.site
Specifies the type of site catalog to be used. One can use either a text based or an xml based site catalog. At present the default is XML.
pegasus.catalog.site.file
The location of file to use as a site catalog. If not specified, then default value of $PWD/sites.xml is used in case of the xml based site catalog.
pegasus.data.configuration

This property sets up Pegasus to run in different environments. This can be set to

sharedfs If this is set, Pegasus will be setup to execute jobs on the shared filesystem on the execution site. This assumes, that the head node of a cluster and the worker nodes share a filesystem. The staging site in this case is the same as the execution site.

nonsharedfs If this is set, Pegasus will be setup to execute jobs on an execution site without relying on a shared filesystem between the head node and the worker nodes.

condorio If this is set, Pegasus will be setup to run jobs in a pure condor pool, with the nodes not sharing a filesystem. Data is staged to the compute nodes from the submit host using Condor File IO.

pegasus.code.generator
The code generator to use. By default, Condor submit files are generated for the executable workflow. Setting to Shell results in Pegasus generating a shell script that can be executed on the submit host.

Files

$PEGASUS_HOME/etc/dax-3.3.xsd
is the suggested location of the latest DAX schema to produce DAX output.
$PEGASUS_HOME/etc/sc-4.0.xsd
is the suggested location of the latest Site Catalog schema that is used to create the XML version of the site catalog
$PEGASUS_HOME/etc/tc.data.text
is the suggested location for the file corresponding to the Transformation Catalog.
$PEGASUS_HOME/etc/sites.xml4 | $PEGASUS_HOME/etc/sites.xml3
is the suggested location for the file containing the site information.
$PEGASUS_HOME/lib/pegasus.jar
contains all compiled Java bytecode to run the Pegasus Workflow Planner.

See Also

pegasus-run(1), pegasus-status(1), pegasus-remove(1), pegasus-rc-client(1), pegasus-analyzer(1)

Authors

Karan Vahi <vahi at isi dot edu>

Pegasus Team http://pegasus.isi.edu