Batch Resources

This document describes the batch resources available at the LSST Data Facility during the interim period where the Rubin filesystems at SLAC go into production mode on our own hardware, and while user and project data are being transferred from NCSA.

Use of SDF batch is documented at

Rubin currently has an allocation of 2000 cores. Currently only a single slurm partition is defined, called “roma”. No special batch queues exist yet.

Running LSST Pipelines with BPS

The LSST Batch Processing Service (BPS) is the standard execution framework for running LSST pipelines using batch resources. There are a few different plugins to BPS that are available that can be used for running BPS on various computing systems:


This section describes how to obtain a personal HTCondor pool and run BPS workflows with it. HTCondor is not included in the LSST stack, so install it locally. Try to import it from python shell first (just in case). If it is not there, install it with: pip3 install --user htcondor Download package condor.tar.gz , and unpack it in your home directory on Rubin machines at SLAC. Edit files to specify your home directory and username:

  • condor/condor_config: change VAR_FOR_HOME

  • condor_master in

Source (if HTCondor is not running )or (for system variables setup only). These scripts will create catalogs, setup system variables and start personal condor ( Create a glide-in to slurm with: sbatch ~/condor/glidein/ Check the status of the glide-in with squeue -u username If glide-in is running, submit your job with bps submit.


The ctrl_bps_parsl package uses the Parsl parallel programming library to enable running on HPC resources. This plugin can also be configured for running on a single node, such as a laptop, which is useful for testing and development. An earlier version of this plugin was developed by DESC and has been used extensively by DESC at NERSC, CC-IN2P3, and CSD3 for running the LSST Science Pipelines at scale. The ctrl_bps_parsl package README has further details about the history, development, and usage of this plugin. The README also has instructions for installing Parsl for use with the LSST Science Pipelines code.

There are nominally four different site configuration classes in ctrl_bps_parsl that can be used for running BPS jobs on the SLAC S3DF cluster. Here is an example BPS configuration file that illustrates possible configurations for each one:

pipelineYaml: "${DRP_PIPE_DIR}/ingredients/LSSTCam-imSim/DRP.yaml"

wmsServiceClass: lsst.ctrl.bps.parsl.ParslService
computeSite: local

  log_level: INFO

    class: lsst.ctrl.bps.parsl.sites.Local
    cores: 8
    class: lsst.ctrl.bps.parsl.sites.Slurm
    nodes: 2
    walltime: 2:00:00     # This is 2 hours
    cores_per_node: 100
    qos: normal
    scheduler_options: |
      #SBATCH --partition=roma
      #SBATCH --exclusive
    class: lsst.ctrl.bps.parsl.sites.TripleSlurm
    nodes: 1
    cores_per_node: 100
    qos: normal
    small_memory: 2.0     # Units are GB
    medium_memory: 4.0
    large_memory: 8.0
    small_walltime: 10.0   # Units are hours
    medium_walltime: 10.0
    large_walltime: 40.0
    class: lsst.ctrl.bps.parsl.sites.work_queue.LocalSrunWorkQueue
    worker_options: "--memory=480000"   # work_queue expects memory in MB
    nodes_per_block: 10

Different configurations are listed, with user-provided labels, under the site section, and the configuration that’s used in the actual BPS submission is specified in the computeSite field via one of those labels.

Monitoring of the pipetask job progress can be enabled by adding the lines

    monitorEnable: true
    monitorFilename: runinfo/monitoring.db

to the desired site subsection. The monitorFilename field specifies the name of the sqlite3 file into which the Parsl workflow tracking information is written. Parsl has a web-app for displaying the monitoring information, and installation of the packages needed to support that web-app are described in the ctrl_bps_parsl README. This python module provides an example for reading the info from that monitoring database.


As of 2022-09-27, the parsl module and its dependencies are only available at S3DF via the CVMFS distributions of lsst_distrib for weekly w_2022_37 and later. However, the modules needed for Parsl monitoring are not available in the CVMFS distributions. They can be installed in ~/.local with the following commands:

$ source /cvmfs/
$ setup lsst_distrib
$ pip install 'parsl[monitoring]' --user
$ pip uninstall sqlalchemy

The pip uninstall sqlalchemy command is needed since the pip install 'parsl[monitoring]' command installs an earlier version of sqlalchemy that’s incompatible with lsst_distrib.

Notes on each of the example configurations follow (Each class listed below lives in the lsst.ctrl.bps.parsl.sites namespace):


This class should be used for running on a single node. The cores field should be set to the number of cores that will be reserved for running the individual pipetask commands, with one core allocated per pipetask job. For example, a Local configuration can be used in an interactive Slurm session obtained using srun

srun --pty --cpus-per-task=8 --mem-per-cpu=4G --time=01:00:00 --partition=roma bash

Note that the --cpus-per-task matches the number of cores in the local config.


This class uses a generic Slurm site configuration that can, in principle, be used with any Slurm submission system.

In the above example, an allocation of 2 nodes with at least 100 cores per node is requested. Various sbatch options can be passed to slurm via the scheduler_options entry. In the above example, we’ve chosen the roma partition at S3DF and requested exclusive use of the nodes.

The bps submit <bps config yaml> command will have Parsl submit a pilot job request to the Slurm queues, and once the pilot job starts, Parsl will run the pipetask jobs on that allocation. Meanwhile, the bps submit command will continue to run on the user’s command line, outputting various log messages from BPS and Parsl. The Slurm configuration class uses Parsl’s HighThroughputExecutor to manage the job execution on the allocated nodes, assigning one core per pipetask job. An important caveat is that the per-pipetask memory requests provided by the BPS config are ignored, so if the average memory per pipetask exceeds 4GB and all of the cores on a S3DF batch node are running, an out-of-memory error will occur, and the Slurm job will terminate. The TripleSlurm and LocalSrunWorkQueue configuration classes provide ways of handling the per-pipetask memory requests.

A useful feature of this class is that it uses the sbatch --dependency=singleton option to schedule a Slurm job that is able to begin execution as soon as the previous job (with the same job name and user) finishes. This way long running pipelines need not request a single, long (and difficult to schedule) allocation at the outset and can instead use a series of smaller allocations as needed.


This configuration provides three HighThroughputExecutors, each with different memory limits for the pipetask jobs that are run on them. In the above example, each executor assigns the specified memory per core, and accordingly limits the number of available cores for running jobs given the total memory per node. Pipetask jobs that request less than 2GB of memory will be run on the “small” allocation; jobs that request between 2GB and 4GB of memory will be run on the “medium” allocation; and all other jobs will be run on the “large” allocation. Despite the segregation into small, medium, and large memory requests, there is still the risk of jobs that request more than 8GB on average causing the “large” allocation to suffer an out-of-memory error.


The LocalSrunWorkQueue configuration class uses Parsl’s WorkQueueExecutor to manage the resource requests by the individual pipetask jobs. It uses the work_queue module to keep track of overall resource usage in the allocation and launches jobs when and where the needed resources are available.

In this class, a Parsl LocalProvider manages the resources from within the allocation itself, and so the procedure for running with this class differs from the Slurm-based classes in that the user is responsible for submitting the pilot job using sbatch command and running the bps submit command within the submission script. In the pilot job, one of the nodes serves as the Parsl “submission node” and runs the pipetask jobs on the available nodes (including the submission node) using the Slurm srun command. Here is an example submission script with the sbatch options set to match the work_queue configuration shown above:


#SBATCH --nodes=10
#SBATCH --exclusive
#SBATCH --time=02:00:00

cd <working_dir>
source /cvmfs/
setup lsst_distrib
<other setup commands>
bps submit <bps yaml file>

Since the Parsl-plugin and other processes running on the submission node have their own memory requirements, one should set the memory available per node to a value somewhat smaller than the total memory capacity. This is done with the worker_options: "--memory=480000" option, where memory is in units of MB. This memory limit applies to all of the nodes in the allocation, so for Slurm jobs that request a large number of nodes, e.g., more than ~20, it would be more efficient to set aside a single node on which to run the bps submit command and use the other nodes as “worker” nodes. This can be accomplished by prepending srun to the bps command in the Slurm batch script:

srun bps submit <bps yaml file>

In this case, one should set #SBATCH --nodes=N so that N is one greater than the nodes_per_block value in the BPS config entry.

To use this class, the work_queue module must be installed. That module is available from the cctools toolkit, which is itself available from conda-forge.