HPC Clusters

vEcoli uses Nextflow and Apptainer containers to run on high-performance computing (HPC) clusters. For users with access to the Covert Lab’s partition on Sherlock, follow the instructions in the Sherlock section. For users looking to run the model on other HPC clusters, follow the instructions in the Other Clusters section.

To speed up HPC workflows, vEcoli supports the HyperQueue executor. See HyperQueue for more information.

Sherlock

On Sherlock, once a workflow is started with runscripts.workflow, runscripts/container/build-image.sh builds an Apptainer image with a minimal snapshot of your cloned repository. Nextflow starts containers using this image to run the steps of the workflow. To run or interact with the model outside of runscripts.workflow, start an interactive container by following the steps in Interactive Container.

Setup

Note

The following setup applies to members of the Covert Lab only.

After cloning the model repository to your home directory, add the following lines to your ~/.bash_profile, then close and reopen your SSH connection:

# Load newer Git, Java (for nextflow), and Python
module load system git java/21.0.4 python/3.12.1
# Include shared Nextflow and HyperQueue installations on PATH
export PATH=$PATH:$GROUP_HOME/vEcoli_env

Then, run the following to test your setup:

python3 runscripts/workflow.py --config configs/test_sherlock.json

This will run a small workflow that:

Builds an Apptainer image with a snapshot of your cloned repository.
Runs the ParCa.
Runs one simulation.
Runs the mass fraction analysis.

All output files will be saved to a test_sherlock directory in your cloned repository. You can modify the workflow output directory by changing the out_dir option under emitter_arg in the config JSON. See Configuration for a description of the Sherlock-specific configuration options and Running Workflows for details about running a workflow on Sherlock.

To run scripts on Sherlock outside a workflow, see Interactive Container. To run scripts on Sherlock through a SLURM batch script, see Non-Interactive Container.

Note

The above setup is sufficient to run workflows on Sherlock. However, if you have a compelling reason to update the shared Nextflow or HyperQueue binaries, navigate to $GROUP_HOME/vEcoli_env and run:

Nextflow: NXF_EDGE=1 nextflow self-update
HyperQueue: See HyperQueue.

Configuration

To tell vEcoli that you are running on Sherlock, you MUST include the following keys in your configuration JSON (note the top-level sherlock key):

{
  "sherlock": {
    # Boolean, whether to build a fresh Apptainer image. If files that are
    # not excluded by .dockerignore did not change since your last build,
    # you can set this to false to skip building the image.
    "build_image": true,
    # Path (relative or absolute, including file name) of Apptainer image to
    # build (or use directly, if build_image is false)
    "container_image": "",
    # Boolean, whether to run using HyperQueue.
    "hyperqueue": false,
    # Boolean, denotes that a workflow is being run as part of Jenkins
    # continuous integration testing. Randomizes the initial seed and
    # ensures that all STDOUT and STDERR is piped to the launching process
    # so they can be reported by GitHub
    "jenkins": false
  }
}

In addition to these options, you MUST set the emitter output directory (see description of emitter_arg in JSON Config Files) to a path with enough space to store your workflow outputs. We recommend setting this to a location in your $SCRATCH directory (e.g. /scratch/users/{username}/out).

If using the Parquet emitter and threaded is not set to false under emitter_arg, a warning will be printed suggesting that you set threaded to false. This ensures that simulations use only a single CPU core, the default that is allocated per simulation on Sherlock (regardless of whether HyperQueue is used). On Sherlock, storage speed is not a bottleneck, so performance with threaded set to false and 1 core per simulation is comparable to running with threaded unset (default: true) and 2 cores per simulation.

If storage speed was a bottleneck, the additional thread would allow simulation execution to continue while Polars writes Parquet files to disk. To properly take advantage of this, you would also need to increase the number of cores per simulation to 2 by modifying the cpus directive under the sherlock and sherlock_hq profiles in runscripts/nextflow/config.template.

Warning

~ and environment variables like $SCRATCH are not expanded in the configuration JSON. See the warning box at Workflows.

Running Workflows

With these options in the configuration JSON, a workflow can be started by running python3 runscripts/workflow.py --config {}, substituting in the path to your config JSON.

Warning

Remember to use python3 to start workflows instead of python.

This command should be run on a login node (no need to request a compute node). If build_image is true in your config JSON, the terminal will report that a SLURM job was submitted to build the container image. When the image build job starts, the terminal will report the build progress.

Note

Files that match the patterns in .dockerignore are excluded from the image.

Warning

Do not make any changes to your cloned repository or close your SSH connection until the build has finished.

Once the build has finished, the terminal will report that a SLURM job was submitted for the Nextflow workflow orchestrator before exiting back to the shell. At this point, you are free to close your connection, start additional workflows, etc. Unlike workflows run locally, Sherlock’s containerized workflows mean any changes made to the repository after the container image has been built will not affect the running workflow.

Once started, the Nextflow job will stay alive for the duration of the workflow (up to 7 days) and submit new SLURM jobs as needed.

If you are trying to run a workflow that takes longer than 7 days, you can use the resume functionality (see Fault Tolerance). Alternatively, consider running your workflow on Google Cloud, which has no maximum workflow runtime (see Google Cloud).

You can start additional, concurrent workflows that each build a new image with different modifications to the cloned repository. However, if possible, we recommend designing your code to accept options through the config JSON which modify the behavior of your workflow without modifying core code. This allows you to save time by reusing a previously built image as follows: set build_image to false and container_image to the path of said image.

There is a 4 hour time limit on each job in the workflow, including analyses. This is a generous limit designed to accomodate very slow-dividing cells. Generally, we recommend that users exclude analysis scripts which take more than a few minutes from their workflow configuration. Instead, either run these manually following Interactive Container or create a SLURM batch script to run these analyses following Non-Interactive Container.

Interactive Container

Warning

The following steps should be run on a compute node. See the Sherlock documentation for details.

The maximum resource request for an interactive compute node is 2 hours, 4 CPU cores, and 8GB RAM/core. Scripts that require more resources should be submitted as SLURM batch scripts to the mcovert or owners partition (see Non-Interactive Container).

To run scripts on Sherlock, you must have either:

Previously run a workflow on Sherlock and have access to the built container image
Built a container image manually using runscripts/container/build-image.sh with the -a flag

Start an interactive container with your full image path (see the warning box at Workflows) by navigating to your cloned repository and running:

runscripts/container/interactive.sh -i container_image -a

Note

Inside the interactive container, you can safely use python directly in addition to the usual uv commands.

The above command launches a container containing a snapshot of your cloned repository as it was when the image was built. This snapshot is located at /vEcoli inside the container and is mostly intended to guarantee reproducibility for troubleshooting failed workflow jobs. More specifically, users who wish to debug a failed workflow job should:

Start an interactive container with the image used to run the workflow.
Use nano to add breakpoints (import ipdb; ipdb.set_trace()) to the relevant scripts in /vEcoli.
Navigate to the working directory (see Troubleshooting) for the job that you want to debug.
Invoke bash .command.sh to run the failing task and pause upon reaching your breakpoints, allowing you to inspect variables and step through the code.

Warning

~ and environment variables like $SCRATCH do not work inside the container. Follow the instructions in the warning box at Workflows outside the container to get the full path to use inside the container.

Danger

Any changes that you make to /vEcoli inside the container are discarded when the container terminates.

To start an interactive container that reflects the current state of your cloned repository, navigate to your cloned repository and run the above command with the -d flag to start a “development” container:

runscripts/container/interactive.sh -i container_image -a -d

In this mode, instead of editing source files in /vEcoli, you can directly edit the source files in your cloned repository and have those changes immediately reflected when running those scripts inside the container. Because you are just modifying your cloned repository, any code changes you make will persist after the container terminates and can be tracked using Git version control.

Non-Interactive Container

To run any script inside a container without starting an interactive session, use the same command as Interactive Container but specify a command using the -c flag. For example, to run the ParCa process, navigate to your cloned repository and run the following command, replacing container_image with the path to your container image and {} with the path to your configuration JSON:

runscripts/container/interactive.sh -i container_image -c "python /vEcoli/runscripts/parca.py --config {}"

This feature is intended for use in SLURM batch scripts to manually run analysis scripts with custom resource requests. Make sure to include one of the following directives at the top of your script:

#SBATCH --partition=owners: This is the largest partition on Sherlock and the most likely to have free resources available for job scheduling. Even so, queue times are variable, and other users may preempt your job at any moment, though this is anecdotally rare for small jobs under an hour long.
#SBATCH --partition=mcovert: Best for high priority scripts (short queue time) that you cannot risk being preempted. The number of available cores is 32 minus whatever is currently being used by other users in the mcovert partition. Importantly, if all 32 cores are in use by mcovert users, not only will your script have to wait for resources to free up, so will any workflows. As such, treat this partition as a limited resource reserved for high priority jobs.
#SBATCH --partition=normal: Potentially longer queue time than either of the two options above but no risk of preemption.
#SBATCH --partition=owners,normal: Uses either the owners or normal partition. This is the recommended option for the vast majority of scripts.

Just as with interactive containers, to run scripts directly from your cloned repository and not the snapshot, add the -d flag and drop the /vEcoli/ prefix from script names. Note that changing files in your cloned repository may affect SLURM batch jobs submitted with this flag.

Other Clusters

Nextflow has support for a wide array of HPC schedulers. If your HPC cluster uses a supported scheduler, you can likely run vEcoli on it with fairly minimal modifications.

Prerequisites

The following are required:

Nextflow (requires Java)
Python 3.9+
Git clone vEcoli to a location that is accessible from all nodes in your cluster
out_dir under emitter_arg set to a location that is accessible from all nodes in your cluster

If your cluster has Apptainer (formerly known as Singularity) installed, check to see if it is configured to automatically mount the filesystem of out_dir (e.g. $SCRATCH). If not, you will need to add -B /full/path/to/out_dir to the containerOptions directives in runscripts/nextflow/config.template, substituting in the absolute path to out_dir. Additionally, you will need to manually specify the same paths when running interactive containers (see Interactive Container) using the -p option.

If your cluster does not have Apptainer, you can try the following steps:

Completely follow the local setup instructions in the README (install uv, etc).
Delete the following lines from runscripts/nextflow/config.template:

process.container = 'IMAGE_NAME'
...
apptainer.enabled = true

Make sure to always set build_runtime_image to false in your config JSONs (see Configuration)

Cluster Options

If your HPC cluster uses the SLURM scheduler, you can use vEcoli on that cluster by changing the queue option in runscripts/nextflow/config.template and runscripts/nextflow/template.nf and all instances of --partition=QUEUE(S) in runscripts.workflow to the right queue(s) for your cluster. Also, remove the --prefer="CPU_GEN... clusterOptions in those same files.

If your HPC cluster uses a different scheduler, refer to the Nextflow executor documentation for more information on configuring the right executor. Beyond the changes above, you will at least need to modify the executor directives for the sherlock and sherlock_hq profiles in runscripts/nextflow/config.template and for the hqWorker process in runscripts/nextflow/template.nf. Additionally, you will need to replace the SLURM submission directives in runscripts.workflow.main() with equivalent directives for your scheduler.

HyperQueue

HyperQueue consists of a head server and one or more workers allocated by the underlying HPC scheduler. By configuring the worker jobs to persist for long enough to complete multiple tasks, HyperQueue reduces the amount of time spent waiting in the queue, which is especially important for workflows with numerous shorter tasks like ours. We recommend using HyperQueue for all workflows that span more than a handful of generations.

Internal Logic

If the hyperqueue option is set to true under the sherlock key in the configuration JSON used to run runscripts/workflow.py, the following steps will occur in order:

If build_image is True, submit a SLURM job to build the container image.
Submit a single long-running SLURM job on the dedicated Covert Lab partition to run both Nextflow and the HyperQueue head server.
Start the HyperQueue head server (initially no workers).
Nextflow submits a SLURM job to run the ParCa then another to create variants. Both must finish for Nextflow to calculate the maximum number of concurrent simulations # seeds * # variants.
Nextflow submits SLURM jobs to start (# seeds * # variants) // 4 HyperQueue workers, each worker with 4 cores, 16GB RAM, and a 24 hour limit. A proportionally smaller worker is potentially created to handle the remainder (e.g. for 2 leftover, 2 cores, 8GB RAM, and same 24 hour limit).
Nextflow submits simulation tasks to the HyperQueue head server, which schedules them on the available workers.
Nextflow submits analysis tasks to SLURM directly as they do not hold up the workflow and can wait for a bit in the queue. This increases simulation throughput by dedicating all HyperQueue worker resources to running simulations.
If any HyperQueue worker job terminates with one of three exit codes (see Fault Tolerance), it is resubmitted by Nextflow to maintain the optimal number of workers for parallelizing the workflow.
As lineages fail and/or complete, the number of concurrent simulations decreases and HyperQueue workers start to go idle. Idle workers automatically terminate after 5 minutes of inactivity.
Upon completion of the Nextflow workflow, the HyperQueue head server terminates any remaining workers and exits.

Monitoring

As long as --server-dir is given as described below, the hq command can be run on any node to monitor the status of the HyperQueue workers and jobs for a given workflow (cheatsheet).

# Replace OUTDIR with the output directory and EXPERIMENT_ID with the
# experiment ID from your configuration JSON.

# Get HyperQueue JOB_ID from this list of jobs
hq --server-dir OUTDIR/EXPERIMENT_ID/nextflow/.hq-server job list

# Get more detailed information about a specific job by ID, including
# its work directory, runtime, and node
hq --server-dir OUTDIR/EXPERIMENT_ID/nextflow/.hq-server job info JOB_ID

Updating

HyperQueue is distributed as a pre-built binary on GitHub. Unfortunately, this binary is built with a newer version of GLIBC than the one available on Sherlock, necessitating a rebuild from source. A binary built in this way is available in $GROUP_HOME/vEcoli_env to users with access to the Covert Lab’s partition on Sherlock. This is added to PATH in the Sherlock setup instructions, so no further action is required.

Users who want or need to build from source should follow these instructions.