Skip to content

Latest commit

 

History

History
138 lines (101 loc) · 4.38 KB

README.md

File metadata and controls

138 lines (101 loc) · 4.38 KB
Raw

launch

Programs for managing job submission through slurm.

These scripts can be used to submit jobs for a cluster running the slurm scheduler. Some scripts are designed with neuroimaging data in mind, for example running a single command to process data from multiple participants in an experiment can be done using slaunch.

Installation

To install the most recent release from PyPI, run:

pip install ezlaunch

To install the latest version on GitHub, run:

pip install git+https://github.com/prestonlab/launch

Verify installation by running launch -h.

Configuration

The launch programs can use information about cluster configuration to fill in blanks (e.g., the number of cores available on each node) and to make sure you're not going over limits of the queue you're submitting to (e.g., a maximum number of nodes).

The configuration is specified using TOML files, which have a simple format. Global settings should go under a [global] heading, while partition-specific settings should go under a heading with the name of that partition. For example (here, "normal" and "development" are two partitions):

[global]
account = "myAccountName"

[normal]
cores = 56
max-nodes = 1280

[development]
cores = 56
max-nodes = 40

This sets the default account to use when submitting jobs. It also sets the number of cores and the maximum number of nodes per job for the normal and development partitions. You may also set a max-cores option to restrict the number of total cores used by a job.

The launch programs will look for a configuration file at $HOME/.launch.toml. If you set the LAUNCH_CONFIG environment variable, launch will look there instead.

EZ Launch

In contrast to launch, which is relatively low level, ezlaunch manages command, batch script, and output files automatically. To use ezlaunch, set the BATCHDIR environment variable to indicate the directory in which job commands, slurm options, and output should be saved.

For example, you could submit a job to a slurm cluster using something like:

ezlaunch -J myjob 'echo "hello world"' -N 1 -n 1 -r 00:10:00

After the job runs, you would have the following files:

$BATCHDIR/myjob1.sh     # job commands 
$BATCHDIR/myjob1.slurm  # slurm options (run time, number of nodes, etc.)
$BATCHDIR/myjob1.out    # output from all commands

If you run another job named "myjob", that will be saved under "myjob2.sh", etc.

Running a command for multiple subjects

The slaunch command makes it quick to process data from multiple subjects in parallel. As a toy example, we can run commands to greet each individual subject:

slaunch -J greet 'echo "hello sub-{}"' 101:102:103 -N 1 -n 3 -r 00:10:00

This will generate three commands corresponding to the three subjects (101, 102, and 103) and run the commands in parallel on different cores. Any {} in the command will be replaced with the subject ID. Subject IDs must be in a list that is separated by colons (:). Optionally, the subject list can be assigned to an environment variable for easier access when running slaunch:

SUBJIDS=101:102:103:104:107:108:110:111
slaunch 'script1 {}' $SUBJIDS -N 1 -n 4 -r 01:00:00
slaunch 'script2 {}' $SUBJIDS -N 1 -n 4 -r 02:00:00

Note that the -n flag doesn't have to match the number of subjects. Here, the 8 subjects will be run in two batches, 4 subjects at a time.

The slaunch command can optionally use the SUBJIDFORMAT environment variable. This allows automatic conversion between subject numbers (e.g., 5) and subject IDs (e.g., sub-05). For that example, you would set:

export SUBJIDFORMAT='sub-%02d'

Then subject number lists can be converted, for example subjids 1:2:3 will produce sub-01:sub-02:sub-03. Use the -g flag when running slaunch to use generated IDs instead of the raw IDs.

Running subjects and runs in parallel

Similarly to slaunch, the rlaunch command can be used to run combinations of subjects and runs in parallel. Here, instead of {}, you use {s} and {r} to indicate where the subject and run numbers, respectively, should be substituted in the run command. For example:

rlaunch 'myscript {s} {r}' 101:102:103 1:2:3:4 -N 1 -n 12 -r 02:00:00

This will generate 12 commands with all combinations of subjects and runs and run those commands in parallel.

Authors

  • Neal Morton
  • Russ Poldrack - wrote original launch script