R: *[Experimental]* Create a launcher with SLURM workers.

crew_launcher_slurm {crew.cluster}

R Documentation

Create a launcher with SLURM workers.

Description

Create an R6 object to launch and maintain workers as SLURM jobs.

Usage

crew_launcher_slurm(
  name = NULL,
  seconds_interval = 0.5,
  seconds_timeout = 60,
  seconds_launch = 86400,
  seconds_idle = Inf,
  seconds_wall = Inf,
  tasks_max = Inf,
  tasks_timers = 0L,
  reset_globals = TRUE,
  reset_packages = FALSE,
  reset_options = FALSE,
  garbage_collection = FALSE,
  launch_max = 5L,
  tls = crew::crew_tls(mode = "automatic"),
  verbose = FALSE,
  command_submit = as.character(Sys.which("sbatch")),
  command_terminate = as.character(Sys.which("scancel")),
  command_delete = NULL,
  script_directory = tempdir(),
  script_lines = character(0L),
  slurm_log_output = "/dev/null",
  slurm_log_error = "/dev/null",
  slurm_memory_gigabytes_per_cpu = NULL,
  slurm_cpus_per_task = NULL,
  slurm_time_minutes = 1440,
  slurm_partition = NULL
)

Arguments

`name`	Name of the launcher.
`seconds_interval`	Number of seconds between polling intervals waiting for certain internal synchronous operations to complete, such as checking `mirai::status()`.
`seconds_timeout`	Number of seconds until timing out while waiting for certain synchronous operations to complete, such as checking `mirai::status()`.
`seconds_launch`	Seconds of startup time to allow. A worker is unconditionally assumed to be alive from the moment of its launch until `seconds_launch` seconds later. After `seconds_launch` seconds, the worker is only considered alive if it is actively connected to its assign websocket.
`seconds_idle`	Maximum number of seconds that a worker can idle since the completion of the last task. If exceeded, the worker exits. But the timer does not launch until `tasks_timers` tasks have completed. See the `idletime` argument of `mirai::daemon()`. `crew` does not excel with perfectly transient workers because it does not micromanage the assignment of tasks to workers, so please allow enough idle time for a new worker to be delegated a new task.
`seconds_wall`	Soft wall time in seconds. The timer does not launch until `tasks_timers` tasks have completed. See the `walltime` argument of `mirai::daemon()`.
`tasks_max`	Maximum number of tasks that a worker will do before exiting. See the `maxtasks` argument of `mirai::daemon()`. `crew` does not excel with perfectly transient workers because it does not micromanage the assignment of tasks to workers, it is recommended to set `tasks_max` to a value greater than 1.
`tasks_timers`	Number of tasks to do before activating the timers for `seconds_idle` and `seconds_wall`. See the `timerstart` argument of `mirai::daemon()`.
`reset_globals`	`TRUE` to reset global environment variables between tasks, `FALSE` to leave them alone.
`reset_packages`	`TRUE` to unload any packages loaded during a task (runs between each task), `FALSE` to leave packages alone.
`reset_options`	`TRUE` to reset global options to their original state between each task, `FALSE` otherwise. It is recommended to only set `reset_options = TRUE` if `reset_packages` is also `TRUE` because packages sometimes rely on options they set at loading time.
`garbage_collection`	`TRUE` to run garbage collection between tasks, `FALSE` to skip.
`launch_max`	Positive integer of length 1, maximum allowed consecutive launch attempts which do not complete any tasks. Enforced on a worker-by-worker basis. The futile launch count resets to back 0 for each worker that completes a task. It is recommended to set `launch_max` above 0 because sometimes workers are unproductive under perfectly ordinary circumstances. But `launch_max` should still be small enough to detect errors in the underlying platform.
`tls`	A TLS configuration object from `crew_tls()`.
`verbose`	Logical, whether to see console output and error messages when submitting worker.
`command_submit`	Character of length 1, file path to the executable to submit a worker job.
`command_terminate`	Character of length 1, file path to the executable to terminate a worker job. Set to `""` to skip manually terminating the worker. Unless there is an issue with the platform, the job should still exit thanks to the NNG-powered network programming capabilities of `mirai`. Still, if you set `command_terminate = ""`, you are assuming extra responsibility for manually monitoring your jobs on the cluster and manually terminating jobs as appropriate.
`command_delete`	Deprecated on 2024-01-08 (version 0.1.4.9001). Use `command_terminate` instead.
`script_directory`	Character of length 1, directory path to the job scripts. Just before each job submission, a job script is created in this folder. Script base names are unique to each launcher and worker, and the launcher deletes the script when the worker is manually terminated. `tempdir()` is the default, but it might not work for some systems. `tools::R_user_dir("crew.cluster", which = "cache")` is another reasonable choice.
`script_lines`	Optional character vector of additional lines to be added to the job script just after the more common flags. An example would be `script_lines = "module load R"` if your cluster supports R through an environment module.
`slurm_log_output`	Character of length 1, file pattern to control the locations of the SLURM worker log files. By default, both standard output and standard error go to the same file. `slurm_log_output = "crew_log_%A.txt"` translates to a line of `⁠#SBATCH --output=crew_log_%A.txt⁠` in the SLURM job script, where `⁠%A⁠` is replaced by the job ID of the worker. The default is `⁠/dev/null⁠` to omit these logs. Set `slurm_log_output = NULL` to omit this line from the job script.
`slurm_log_error`	Character of length 1, file pattern for standard error. `slurm_log_error = "crew_log_%A.txt"` translates to a line of `⁠#SBATCH --error=crew_log_%A.txt⁠` in the SLURM job script, where `⁠%A⁠` is replaced by the job ID of the worker. The default is `⁠/dev/null⁠` to omit these logs. Set `slurm_log_error = NULL` to omit this line from the job script.
`slurm_memory_gigabytes_per_cpu`	Positive numeric of length 1 with the gigabytes of memory required per CPU. `slurm_memory_gigabytes_per_cpu = 2.40123` translates to a line of `⁠#SBATCH --mem-per-cpu=2041M⁠` in the SLURM job script. `slurm_memory_gigabytes_per_cpu = NULL` omits this line.
`slurm_cpus_per_task`	Optional positive integer of length 1, number of CPUs for the worker. `slurm_cpus_per_task = 4` translates to a line of `⁠#SBATCH --cpus-per-task=4⁠` in the SLURM job script. `slurm_cpus_per_task = NULL` omits this line.
`slurm_time_minutes`	Numeric of length 1, number of minutes to designate as the wall time of `crew` each worker instance on the SLURM cluster. `slurm_time_minutes = 60` translates to a line of `⁠#SBATCH --time=60⁠` in the SLURM job script. `slurm_time_minutes = NULL` omits this line.
`slurm_partition`	Character of length 1, name of the SLURM partition to create workers on. `slurm_partition = "partition1,partition2"` translates to a line of `⁠#SBATCH --partition=partition1,partition2⁠` in the SLURM job script. `slurm_partition = NULL` omits this line.

Details

WARNING: the crew.cluster SLURM plugin is experimental and has not actually been tested on a SLURM cluster. Please proceed with caution and report bugs to https://github.com/wlandau/crew.cluster.

To launch a SLURM worker, this launcher creates a temporary job script with a call to crew::crew_worker() and submits it as an SLURM job with sbatch. To see most of the lines of the job script in advance, use the script() method of the launcher. It has all the lines except for the job name and the call to crew::crew_worker(), both of which will be inserted at the last minute when it is time to actually launch a worker.

Attribution