Using SLURM Defaults in parade

What are defaults and why use them?

The parade package allows you to set site/project defaults for SLURM job submission, making your R code portable across different clusters and reducing repetitive resource specifications. Instead of specifying the same SLURM parameters (partition, memory, CPU count, etc.) in every job submission, you can configure them once and reuse them throughout your project.

Key benefits of using defaults:

• Portability: Your R scripts work across different clusters without modification
• Consistency: Ensure all jobs use appropriate resource allocations for your environment
  
• Flexibility: Selectively omit problematic flags that some clusters reject (e.g., --mem)
• Maintainability: Update resource requirements in one place rather than throughout your codebase

Defaults are stored in a JSON configuration file and are accessible from R through simple functions.

Quick start

library(parade)
paths_init()

# Set up defaults for your cluster environment
slurm_defaults_set(
  partition = "general",     # Default partition to use
  time = "2h",              # Default time limit
  cpus_per_task = 16,       # Default CPU allocation
  mem = NA,                 # <- omit --mem entirely (some clusters reject this flag)
  omp_num_threads = 1,      # Default OpenMP thread count
  persist = TRUE            # Save these defaults to config file
)

# Optionally set a default template path
slurm_template_set("registry://templates/parade-slurm.tmpl")

# Now submit jobs using your defaults
job <- submit_slurm("script.R")

When you call submit_slurm(), it automatically merges your call’s resources parameter with the configured defaults, normalizes time specifications, and drops any NA or omit() fields before constructing the #SBATCH directives for your SLURM job.

Note for flows: to run a flow on SLURM with your defaults, build a resource set with slurm_resources() and pass it via dist_slurm() inside distribute(), e.g.:

flow(grid) |>
  stage("analyze", analyze_fn) |>
  distribute(dist_slurm(resources = slurm_resources(profile = "standard")))

Configuration file

Parade searches for configuration files in the following order:

1. PARADE_CONFIG environment variable (exact file path), if set
2. <project>/parade.json, if present
   
3. <project>/.parade/parade.json (created automatically as needed)

The configuration file uses JSON format. Here’s an example showing typical defaults for a research cluster:

{
  "slurm": {
    "template": "registry://templates/parade-slurm.tmpl",
    "defaults": {
      "partition": "general",
      "time": "2h", 
      "cpus_per_task": 16,
      "mem": null,
      "omp_num_threads": 1
    }
  }
}

Note that null in JSON corresponds to NA in R, which tells parade to omit that parameter entirely from the SLURM submission.

Working with defaults programmatically

Inspecting current defaults

# View all current defaults
slurm_defaults_get()

# Check specific default values
defaults <- slurm_defaults_get()
defaults$time        # "2h"
defaults$partition   # "general"

Building resource lists with overrides

The slurm_resources() function combines defaults with job-specific overrides:

# Use defaults but override time and explicitly omit memory
resources <- slurm_resources(list(time = "90min", mem = omit()))

# Submit job with these specific resources
job <- submit_slurm("analysis.R", resources = resources)

Updating defaults during your session

# Change defaults temporarily (session only)
slurm_defaults_set(mem = NA)           # omit --mem flag
slurm_defaults_set(cpus_per_task = 8)  # reduce CPU count

# Make changes permanent by saving to config file
slurm_defaults_set(time = "1h", persist = TRUE)

# Set multiple defaults at once
slurm_defaults_set(
  partition = "gpu",
  time = "4h", 
  gres = "gpu:1",
  persist = TRUE
)

Overriding defaults in job submission

You can override defaults on a per-job basis by passing a resources argument to submit_slurm():

# Use defaults for most parameters, but need more time and memory for this job
big_job <- submit_slurm("big_analysis.R", 
                        resources = list(time = "12h", mem = "32G"))

# For a quick test job, use minimal resources
test_job <- submit_slurm("test.R", 
                         resources = list(time = "5min", cpus_per_task = 1))

# Submit to a different partition while keeping other defaults
gpu_job <- submit_slurm("model_training.R",
                        resources = list(partition = "gpu", gres = "gpu:2"))

Common use cases and examples

Setting up defaults for different cluster environments

For a cluster that rejects memory specifications:

slurm_defaults_set(
  partition = "compute",
  time = "2h",
  cpus_per_task = 16,
  mem = NA,              # Omit memory specification
  persist = TRUE
)

For a GPU cluster:

slurm_defaults_set(
  partition = "gpu",
  time = "4h", 
  cpus_per_task = 8,
  mem = "16G",
  gres = "gpu:1",
  persist = TRUE
)

For high-memory jobs:

slurm_defaults_set(
  partition = "highmem",
  time = "8h",
  cpus_per_task = 32,
  mem = "128G",
  persist = TRUE
)

Best practices

1. Set defaults early: Configure defaults at the beginning of your project after initializing with paths_init()

2. Use meaningful time limits: Set reasonable default time limits to avoid jobs sitting in queue unnecessarily

3. Consider cluster policies: Some clusters reject certain flags (like --mem) - use NA to omit these

4. Environment-specific configs: Use different config files or the PARADE_CONFIG environment variable for different clusters

5. Version control: Consider committing your parade.json file to version control so team members share the same defaults

This approach keeps your R code portable across clusters with different SLURM policies and resource requirements.