crew.cluster: crew launcher plugins for traditional high-performance computing clusters

In computationally demanding analysis projects, statisticians and data scientists asynchronously deploy long-running tasks to distributed systems, ranging from traditional clusters to cloud services. The crew.cluster package extends the mirai-powered ‘crew’ package with worker launcher plugins for traditional high-performance computing systems. Inspiration also comes from packages mirai, future, rrq, clustermq, and batchtools.

Installation

Documentation

Please see https://wlandau.github.io/crew.cluster/ for documentation, including a full function reference and usage tutorial.

Usage

First, create a controller object appropriate for your platform. For example, to launch workers on a Sun Grid Engine (SGE) cluster, use crew_controller_sge().

library(crew.cluster)
controller <- crew_controller_sge(
  name = "my_workflow", # for informative job names
  workers = 16,
  tasks_max = 2, # to avoid reaching wall time limits
  seconds_idle = 10, # to release resources when they are not needed
  script_lines = "module load R" # if R is an environment module
)
controller$start()

At this point, usage is exactly the same as basic crew. The push() method submits tasks and auto-scales SGE workers to meet demand.

controller$push(name = "do work", command = do_work())

The pop() method retrieves available tasks.

controller$pop()
#> # A tibble: 1 × 11
#>   name         command result seconds   seed error trace warni…¹ launc…² worker insta…³
#>   <chr>        <chr>   <list>   <dbl>  <int> <chr> <chr> <chr>   <chr>    <int> <chr>  
#> 1 do work   … do_work… <int>        0 1.56e8 NA    NA    NA      79e71c…      1 7686b2…
#> # … with abbreviated variable names ¹​warnings, ²​launcher, ³​instance

Remember to terminate the controller when you are done.

controller$terminate()

Monitoring

To manage resource usage, you may choose to list and manually terminate cluster jobs using crew_monitor_sge() and other supported monitors. Example for SGE:

monitor <- crew_monitor_sge()
job_list <- monitor$jobs()
job_list
#> # A tibble: 2 × 9
#>   job_number prio    name    owner state start_time queue_name jclass_name slots
#>   <chr>      <chr>   <chr>   <chr> <chr> <chr>      <chr>      <lgl>       <chr>
#> 1 131853812  0.05000 crew-m… USER… r     2024-01-0… all.norma… NA          1    
#> 2 131853813  0.05000 crew-m… USER… r     2024-01-0… all.norma… NA          1
monitor$terminate(jobs = job_list$job_number)
#> USER has registered the job 131853812 for deletion
#> USER has registered the job 131853813 for deletion
monitor$jobs()
#> data frame with 0 columns and 0 rows

monitor$terminate(all = TRUE) terminates all your SGE jobs, regardless of whether crew.cluster created them.

Tips

• crew.cluster submits jobs over the local network using system calls to the resource manager (e.g SGE or SLURM). Please invoke crew.cluster on a node of the cluster, either a login node (head node) or a compute node.
• Most clusters install software like R in versioned environment modules. Most likely, you will need to call module load R (or module load R/x.y.z for a specific version) in order to use on the cluster. In crew.cluster, you will most likely need to supply "module load R" or similar to the script_lines argument of e.g. crew_controller_sge().

Risks

The risks of crew.cluster are the same as those of crew, plus the risks of traditional high-performance computing environments. These distributed systems typically operate inside a firewall and trust the local network. It is your responsibility to assess the security of these systems and use crew.cluster in a safe manner. In addition, crew.cluster automatically launches jobs on the cluster scheduler, and it may not always be able to terminate leftover jobs. It is your responsibility to monitor your jobs and manually terminate jobs that crew.cluster may not be able to.

Thanks

• Charlie Gao created mirai and nanonext and graciously accommodated the complicated and demanding feature requests that made crew and its ecosystem possible.
• Thanks to Michael Schubert for sharing the template files in clustermq under the permissive Apache License 2.0. These scripts helped construct launcher plugins to clusters where direct access was not possible. See the LICENSE.note file in this package.

Code of Conduct

Please note that the crew project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

Citation

citation("crew.cluster")