How to run the ASKAP pipelines¶

Loading the pipeline module¶

The pipeline scripts are now accessed through a specific module on galaxy - askappipeline. This is separate to the main askapsoft module, to allow more flexibility in updating the pipeline scripts. To use, simply run:

module load askappipeline

This provides access to the main pipeline executable processASKAP.sh.

Some parts of the pipeline make use of other modules, which are loaded at the appropriate time. The beam footprint information is obtained by using the schedblock tool in the askappy module, while beam locations are set using footprint from the same module.

For the case of either BETA data or observations made with a non-standard footprint, the footprint tool will not have the correct information, and the legacy tool footprint.py is used (this is a holdover from the original ACES software repository).

Once loaded, the askappipeline module will set an environment variable $PIPELINEDIR, pointing to the directory containing the scripts. It also defines $PIPELINE_VERSION to be the version number of the currently-used module.

Pipeline configuration¶

The pipeline is configured with a range of input parameters, all of which have some default value. The default is changed by one of the following methods:

a configuration file via:
```
processASKAP.sh -c myInputs.sh
```

a template configuration plus the scheduling-block IDs:

processASKAP.sh -s SB_SCIENCE -b SB_1934 -p SB_PB -t template.sh

or a template configuration and a configuration file specific to the observation being processed:
```
processASKAP.sh -t template.sh -c myInputs.sh
```

Parameters are decided by applying, in order, the pipeline defaults, the template (if given) then the configuration file (so that parameters given the configuration file have precedence over the corresponding value in the template). Scheduling block IDs given on the command-line take precendence over any given in the configuration (or template) files.

The template and configuration files are shell scripts that define environment variables. A configuration file could look something like this:

#!/bin/bash -l
#
# Example user input file for ASKAP processing.
# Define variables here that will control the processing.
# Do not put spaces either side of the equals signs!
# control flags
SUBMIT_JOBS=true
DO_SELFCAL=false
# scheduling blocks for calibrator & data
SB_1934=507
SB_SCIENCE=514
# base names for MS and image data products
MS_BASE_SCIENCE=B1740_10hr.ms
IMAGE_BASE_CONT=i.b1740m517.cont
# other imaging parameters
NUM_PIXELS_CONT=4096
NUM_TAYLOR_TERMS=2
CORES_PER_NODE_CONT_IMAGING=15

This file should define enough environment variables for the scripts to run successfully. Mandatory ones, if you are starting from scratch, are the locations of either the SBs for the observations or the specific MSs.

It is possible for processASKAP.sh to read the template from the SB parset, where it looks for the parameter common.cp.processing_template. If this option is used, the SBID command-line options must be given:

processASKAP.sh -s SB_SCIENCE -b SB_1934

Giving the template on the command-line (with -t) will override the template provided in the SB parset. If the SB parset does not specify a parset, a warning will be given and processing will continue without it (using the defaults and any configuration file given).

There are two other command-line parameters that are related to how the jobs are run. One relates to the QUEUE, which is the name of the slurm partition on which the bulk of the jobs will be run. While the template or config can specify a value for QUEUE, this is overridden by using the -q <QUEUE> option. For example:

processASKAP.sh -s SB_SCIENCE -b SB_1934 -q work

This will send the jobs to the work partition instead of the default (askaprt). The execeptions to this are jobs that require access to the /askapbuffer filesystem - pipeline launch/relaunch jobs, initial raw data access jobs, or submission to CASDA. These are sent to the queue indicated by QUEUE_OPS, which defaults to askaprt, and shouldn’t in general be altered.

The other, -S <version>, provides a singularity version that can be used to override the default value. By default on setonix, jobs that are run on the askaprt partition use a singularity version ending in ‘-askap’, which mounts the /askapbuffer filesystem. Jobs that run on the work partition, which don’t have visibility to that filesystem, use singularity versions ending in ‘-mpi’. Providing a version directly with the -S option will use that version for all jobs, irrespective of the partition they are run on.

When run, the pipeline configuration parameters will be archived in the slurmOutputs directory (see below). It will be called pipelineConfig__<time>.sh, where <time> will be the timestamp of the pipeline run. This file will combine the user configuration and the template configuration, so that you’ll be able to keep a record of exactly what you have run.

Important Note: The input file is a bash script, so formatting matters. Most importantly in this case, you can not have spaces either side of the equals sign when defining a variable.

User-defined Environment Variables¶

The user is able to specify environment variables that directly relate to the parset parameters for the individual ASKAPsoft tasks. The input parameters are named differently, so that they are tied more obviously to specific tasks, and to distinguish between the the same parset parameter used for different jobs (eg. the preconditioning definition could differ for the continuum & spectral-line imaging).

The input environment variables are all given with upper-case names, with an underscore separating words (eg. VARIABLE_NAME). This will distinguish them from the parset parameters.

The following pages list the environment variables defined in defaultConfig.sh – these can all be redefined in your input file to set and tweak the processing. The default value of the parameter (if it has one) is listed in the tables, and in many of the tables the parset parameter than the environment variable maps to is given.

What is created and where does it go?¶

Any measurement sets, images and tables that are created are put in an output directory specified in the input file (if not provided, they go in the directory in which processASKAP.sh is run). There will be a file called PROCESSED_ON that holds the timestamp indicating when the script was run (this timestamp is used in various filenames). Also created are a number of subdirectories which hold various types of files. These are:

slurmFiles/ – the files in here are the job files that are submitted to the queue via the sbatch command. When a job is run, it makes a copy of the file that is labelled with the job ID.
metadata/ – information about the measurement sets and the beam footprint are written to files here. See below for details on files.
parsets/ – any parameter sets used by the askapsoft applications are written here. These contain the actual parameters that are used by the various programs. These are labeled by the job ID.
logs/ – the logs that are written by the askapsoft applications themselves are put here.
slurmOutputs/ – the stdout and stderr from the slurm job itself are written to these files. Such files are usually slurm-XXXXXX.out (XXXXXX being the job ID), but these scripts rename the files so that the filename shows what job relates to what file (as well as providing the ID).
diagnostics/ - this directory is intended to hold plots and other data products that indicate how the processing went. The pipeline only produces a few particular types at the moment, but the intention is this will expand with time.
tools/ – utility scripts to show progress and kill all jobs for a given run are placed here. See Diagnostics and job management for details.
clink/ - yaml files produced for CLINK notifications are written here (see notifications section below)
pipelineControl/ - parsets containing control and check parameters are kept here. See Diagnostics and job management for more details.

There are two additional sets of files that are worth mentioning - the ‘stats’ files, and the pipeline control parsets (found in the pipelineControl directory) - which are used to track progress and monitor resource usage. These are described further on the Diagnostics and job management page.

Metadata directory¶

The metadata directory contains a number of files that hold important metadata describing the observation that is not yet available through the measurement set. These files are used at various stages in the processing. The files are:

mslist-<timestamp><index>.txt - Results of running mslist (mslist (Measurement summary and data inspection utility)) on a measurement set. This summarises the metadata of the observation, including: time range; spectral window (frequency range, number of channels, etc); field list (where the telescope was pointing); list of antennas used. When the observation has more than one raw MS, only one is used (as they should all have the same metadata), and the <index> addition will indicate which file was used. If there is a single MS in the raw directory, there will only be a timestamp. There may be other mslist files created for the intermediate MSs created during the pipeline processing, depending on the modes used.
mslist-cal-<timestamp>.txt - As above, but for the bandpass calibration observation.
mslist-<timestamp><index>.txt.timerange - When the DO_SPLIT_TIMEWISE=true option is used, this lists the boundaries (start and end) of each time range that the data will be split into.
fieldlist-<timestamp><index>.txt - Extracted from the mslist output, this is the list of fields for the science observation. This is used to define the footprint positions.
schedblock-info-<SBID>.txt - Summary of the information from the scheduling block database. For each SBID, this lists the summary information, the parameters and the variables. This includes the specification for the beam footprint, which is used to form the next file.
footprintOutput-sb<SBID>-<FIELD>.txt - A list of the locations and offsets of each beam in the footprint used for a given field of the science observation. Each field (ie. pointing direction) in the MS gets its own footprint file. This is used by the pipeline to define the locations of the individual beam images.

As detailed in User parameters - Archiving, the metadata folder will be copied into each archived MS prior to it being tarred, and also provided in the evaluation files made available in CASDA.

Measurement sets¶

To provide the input data to the scripts, you can provide either the scheduling blocks (SBs) of the two observations, or provide specific measurement sets (MSs) for each case.

The measurement sets presented in the scheduling block directories have a variety of forms:

The earliest observations had all beams in a single measurement set. For these, splitting with mssplit is required to get a single-beam MS used for processing.
Many observations have one beam per MS. If no selection of channels or fields or scans is required, these can be copied rather than split. Note that splitting of the bandpass observations are necessary, to isolate the relevant scan.
Recent large observations have more than one MS per beam, split in frequency chunks. The pipeline will merge these to form a single local MS for each beam. If splitting (of channels, fields or scans) is required, this is done first, before merging the local subsets.

The measurement sets that will be created should be named in the configuration file. A wildcard %s can be used to represent the scheduling block ID, and %b should be used to represent the beam number in the resulting MSs, since the individual beams will be split into separate files.

Each step detailed below can be switched on or off, and those selected will run fine (provided any pre-requisites such as measurement sets or bandpass solutions etc are available). If you have already created an averaged science MS, you can re-use that with the MS_SCIENCE_AVERAGE parameter (see User Parameters - Preparation of the Science field data), again with the %b wildcard to represent the beam number and %s the scheduling block ID.

Workflow summary¶

Here is a summary of the workflow provided for by these scripts:

Get observation metadata from the MS and the beam footprint. This does the following steps:
- Use mslist to get basic metadata for the observation, including number of antennas & channels, and the list of field names. (If merging is required, additional metadata files will be created later.)
- Use schedblock to determine the footprint specification and other observation details.
- Use footprint to convert that into beam centre positions.
Read in user-defined parameters from the provided configuration file, and define further parameters derived from them.
If bandpass calibration is required and a 1934-638 observation is available, we split out the relevant beams with mssplit (mssplit (Measurement Splitting/Averaging Utility)) into individual measurement sets (MSs), one per beam. Merging may be required as described above. Only the scan in which the beam in question was pointing at 1934-638 is used - this assumes the beams were pointed at it in order (so that beam 0 was pointing at in in scan 0, etc)
The local MSs are flagged using cflag (cflag / cflagger (Flagging Utility)) in two passes: first, selection rules covering channels, time ranges, antennas & baselines, and autocorrelations are applied, along with an optional simple flat amplitude threshold; then a second pass that covers Stokes-V and dynamic amplitude flagging, that integrate individual spectra.
There is an option to use the AOFlagger tool instead of cflag to do the flagging, with the ability to provide strategy files for each flagging task.
The bandpass solution is then determined with cbpcalibrator (cbpcalibrator), using all individual MSs and stored in a single CASA table.
If bandpass calibration has already been done in another directory, the bandpass (and leakage) tables and their plots are copied to the BPCAL directory in the current processing area, and thereafter used from there.

The science field is processed for each field name - what follows describes the steps used for each field:

The science field data is split with mssplit, producing one measurement set per beam. You can select particular scans or fields here, but the default is to use everything. Each field gets its own directory. If the data was taken with the file-per-beam mode, and no selection is required, a direct copy is used instead of mssplit. Again, merging with msmerge may be required for some datasets. Any splitting that is needed in that case is done first.
The bandpass solution is then applied to each beam MS with ccalapply (ccalapply (Calibration Applicator)).
Flagging is then applied to the bandpass-calibrated dataset. The same procedure as for the calibrator is used, with separate user parameters to control it.
The science field data are then averaged with mssplit to form continuum data sets. (Still one per beam).
Another round of flagging can be done, this time on the averaged dataset.
Each beam is then imaged individually. This is done in one of two ways:
- Basic imaging with either imager (imager) or cimager (cimager), without any self-calibration. A multi-scale, multi-frequency clean is used, with major & minor cycles.
- With self-calibration. First we image the field with imager or cimager as for the first option. selavy (Selavy Basics) is then used to find bright components, which are then used with ccalibrator (ccalibrator) to calibrate the gains, and we then re-image using the calibration solution. This process is repeated a number of times.
The calibration solution can then be applied directly to the MS using ccalapply, optionally creating a copy in the process.
The continuum dataset can then be optionally imaged as a “continuum cube”, using imager or simager (simager) to preserve the full frequency sampling. This mode can be run for a range of polarisations, creating a cube for each polarisation requested.
Once the continuum image has been made, the source-finder selavy can be run on it to produce a deeper catalogue of sources.
Fast, or transient imaging - that is, imaging in the continuum at a series of narrow time intervals to allow transient source detection - may be applied using both the deep continuum clean model and source catalogue.
The beam images can optionally be smoothed to a common resolution - the smallest one large enough to cover all input beam images’ PSFs. This produces new versions of the restored images, which are used for subsequent source-finding as the primary image.
The beam images (continuum & continuum-cubes) may be convolved to ensure they all have the same PSF.
Once all beams have been done, they are all mosaicked together using linmos-mpi (linmos (Linear Mosaic Applicator)). This is done for all image types (restored, convolved-restored, residual, model image). This applies a primary-beam correction using the nominated holography beam data. Use the logs to find what the beam arrangement for your observation was. After mosaicking, selavy* can be run on the final image to create the final source catalogue.
Additionally, spectral imaging of the full-resolution MS for individual beams can be done. There are several optional steps to further prepare the spectral-line dataset:
- A nominated channel range can be copied to a new MS with mssplit.
- The gains solution from the continuum self-calibration can be applied to the spectral-line MS using ccalapply.
- The continuum can be subtracted from the spectral-line MS (using a model of the selavy catalogue or the clean model from the continuum imaging) using ccontsubtract (ccontsubtract).
Once the spectral-line dataset is prepared, imager or simager is used to do the spectral-line imaging. This creates a cube using a large number of processors, each independently imaging a single channel.
There is a further task to remove any residual continuum from the image cube by fitting a low-order polynomial to each spectrum independently.
Source-finding with selavy can then be run on the spectral-cubes.
A diagnostics script is run to produce QA & related plots. Details of diagnostic plots can be found on Validation and Diagnostics.
Further validation scripts may also be run - these are typically provided by ASKAP Science teams, to produce validation reports that are sent to CASDA. Further details are on Validation and Diagnostics.
Finally, if required and once all the processing is deemed complete, the necessary data products are identified for upload to CASDA, with an interchange file called observation.xml created in a local directory (tagged with the timestamp of the pipeline run).
Any job errors or failures to complete will cause the pipeline to be re-run - if the errors were due to transient platform issues, these will likely succeed on subsequent passes. If it is a persistent error, the pipeline will only run a maximum number of times (10 by default) before stopping. See below for further details.
If all jobs have completed, an additional pass with a different template or config file is possible (by setting SUBSEQUENT_TEMPLATE and/or SUBSEQUENT_CONFIG in the original template/config).
Once all pipeline processing is complete, the observation.xml files for all pipeline runs are merged and sent to the appropriate CASDA area for ingest.

‘Rapid’ mode processing¶

The “rapid survey” mode of the telescope involves taking many relatively short integrations of different fields within a single scheduling block, allowing the coverage of wide areas of sky. A consequence of this mode is that the MS sizes are quite small, so that the pre-imaging tasks (splitting, bandpass calibration, flagging, averaging) take very little time. This can lead to large overheads for the slurm scheduling system, especially when combined with the large numbers of jobs coming from the many fields.

To address this, there is a rapid pipeline mode that combines all these pre-imaging steps into a single script (that is written to the tools/ directory when processASKAP.sh is run). This script is called from within the imaging job itself, so that each beam only needs to run the continuum imaging/self-cal job (as well as the jobs that follow - applying calibration, Stokes V imaging, linmos), greatly reducing the load on the slurm controller.

This mode is turned on by setting DO_RAPID_SURVEY=true. There are some constraints with this:

Only continuum imaging is done in this mode.
There is no merging of spectral chunks of the raw data - it is assumed the raw data is either in a single MS or in one MS per beam.

Otherwise, the processing is exactly the same as for the regular pipeline case.

Pipeline restarts¶

The pipeline makes use of the stats and pipelineControl.in files (see Diagnostics and job management for details) to determine whether there were failed jobs, or jobs that started but did not complete (either by timing out or being cancelled).

A ‘launch’ job is submitted to run after the final pipeline job, and this job determines if there were failures in the previous run. If so, processASKAP.sh is run again with the same command-line arguments. The check parameters in pipelineControl.in will allow the new pipeline run to begin only from the jobs not yet complete. If the errors were due to transient compute-node or filesystem issues, then it is likely they will run successfully the next time round.

If the errors are related to something with the data, it is likely the same problem will recur. In this case, there is a limit to the maximum number of resubmissions that can be made, encapsulated by the MAX_RESUBMIT parameter (and stored in the file given by SUBMIT_TRACKER).

The restart can be forced (so that it re-submits the pipeline regardless of whether there were failed jobs) by ensuring the file indicated by FORCE_RESTART_FILE exists (this defaults to .pipeline-start). This file can be empty (creating it with a simple “touch .pipeline-start” is sufficient). It gets removed when the pipeline is restarted. This will also increment the submission counter in SUBMIT_TRACKER.

Notifications of failures & completion¶

When run operationally (as the ‘askapops’ user), the pipeline makes use of the “CLINK” Component Linking system to send out notifications of events such as job failures and pipeline completion. The specific events that trigger a notification are:

Completion of all jobs in a pipeline.
Failure to launch jobs during a processASKAP.sh execution (for instance, due to a slurm error).
The failure of an individual slurm job.
A job that has timed-out and not completed. The notification for this is only sent out at the pipeline restart stage (see previous section).
The restart of a pipeline.

While these notifications are only sent to CLINK when the clink executable is available (which typically means for the askapops user alone), each notification is sent with a yaml file that is created in the clink/ subdirectory. These yaml files are created for all users of the pipeline, and are available for examination.