Parallelize R code on a Slurm cluster

Many computing-intensive processes in R involve the repeated evaluation of a function over many items or parameter sets. These so-called embarrassingly parallel calculations can be run serially with the lapply or Map function, or in parallel on a single machine with mclapply or mcMap (from the parallel package).

The rslurm package simplifies the process of distributing this type of calculation across a computing cluster that uses the Slurm workload manager. Its main function, slurm_apply (and the related slurm_map) automatically divide the computation over multiple nodes and write the necessary submission scripts. The package also includes functions to retrieve and combine the output from different nodes, as well as wrappers for common Slurm commands.

Table of contents

Basic example

To illustrate a typical rslurm workflow, we use a simple function that takes a mean and standard deviation as parameters, generates a million normal deviates and returns the sample mean and standard deviation.

test_func <- function(par_mu, par_sd) {
    samp <- rnorm(10^6, par_mu, par_sd)
    c(s_mu = mean(samp), s_sd = sd(samp))

We then create a parameter data frame where each row is a parameter set and each column matches an argument of the function.

pars <- data.frame(par_mu = 1:10,
                   par_sd = seq(0.1, 1, length.out = 10))
head(pars, 3)
##   par_mu par_sd
## 1      1    0.1
## 2      2    0.2
## 3      3    0.3

We can now pass that function and the parameters data frame to slurm_apply, specifiying the number of cluster nodes to use and the number of CPUs per node. The latter (cpus_per_node) determines how many processes will be forked on each node, as the mc.cores argument of parallel::mcMap.

sjob <- slurm_apply(test_func, pars, jobname = 'test_apply',
                    nodes = 2, cpus_per_node = 2, submit = FALSE)
## Submission scripts output in directory _rslurm_test_apply

The output of slurm_apply is a slurm_job object that stores a few pieces of information (job name, job ID, and the number of nodes) needed to retrieve the job's output.

The default argument submit = TRUE would submit a generated script to the Slurm cluster and print a message confirming the job has been submitted to Slurm, assuming your are running R on a Slurm head node. When working from a R session without direct access to the cluster, you must set submit = FALSE. Either way, the function creates a folder called \_rslurm\_[jobname] in the working directory that contains scripts and data files. This folder may be moved to a Slurm head node, the shell command sbatch run from within the folder, and the folder moved back to your working directory. The contents of the \_rslurm\_[jobname] folder after completion of the test_apply job, i.e. following either manual or automatic (i.e. with submit = TRUE) submission to the cluster, includes one results_*.RDS file for each node:

list.files('_rslurm_test_apply', 'results')
## [1] "results_0.RDS" "results_1.RDS"

The results from all the nodes can be read back into R with the get_slurm_out() function. In this example, wait = FALSE, but if you use the default argument wait = TRUE, execution will be paused until the Slurm job finishes running.

res <- get_slurm_out(sjob, outtype = 'table', wait = FALSE)
head(res, 3)
##       s_mu       s_sd
## 1 1.000137 0.09995552
## 2 2.000144 0.19988175
## 3 2.999822 0.30030102

The utility function print_job_status displays the status of a submitted job (i.e. in queue, running or completed), and cancel_slurm will remove a job from the queue, aborting its execution if necessary. These functions are R wrappers for the Slurm command line functions squeue and scancel, respectively.

When outtype = 'table', the outputs from each function evaluation are row-bound into a single data frame; this is an appropriate format when the function returns a simple vector. The default outtype = 'raw' combines the outputs into a list and can thus handle arbitrarily complex return objects.

res_raw <- get_slurm_out(sjob, outtype = 'raw', wait = FALSE)
## [[1]]
##       s_mu       s_sd 
## 1.00013690 0.09995552 
## [[2]]
##      s_mu      s_sd 
## 2.0001445 0.1998817 
## [[3]]
##     s_mu     s_sd 
## 2.999822 0.300301

The utility function cleanup_files deletes the temporary folder for the specified Slurm job.


Single function evaluation

In addition to slurm_apply, rslurm also defines a slurm_call function, which sends a single function call to the cluster. It is analogous in syntax to the base R function, accepting a function and a named list of parameters as arguments.

sjob <- slurm_call(test_func, jobname = 'test_call',
                   list(par_mu = 5, par_sd = 1), submit = FALSE)
## Submission scripts output in directory _rslurm_test_call

Because slurm_call involves a single process on a single node, it does not recognize the nodes and cpus_per_node arguments; otherwise, it accepts the same additional arguments (detailed in the sections below) as slurm_apply.


Applying a function to a list of complex objects

The function passed to slurm_apply can only receive atomic parameters stored within a data frame. Suppose we want instead to apply a function func to a list of complex R objects, obj_list. In that case we can use the function slurm_map, which is similar in syntax to lapply from base R and map from the purrr package. Its first argument is a list which can contain objects of any type, and its second argument is a function that acts on a single element of the list.

sjob <- slurm_map(obj_list,
                  nodes = 2, cpus_per_node = 2)

The output generated by slurm_map is structured the same way as slurm_apply. The procedures for checking the job status, extracting the results of the job, and cleaning up job files are also the same as described above.

Adding auxiliary data and functions

Each of the tasks started by slurm_apply and slurm_map begin by default in an "empty" R environment containing only the function to be evaluated and its parameters. If we want to pass additional arguments to the function that do not vary with each task, we can simply add them as additional arguments to slurm_apply or slurm_map, like in this example, where we want to take the logarithm of many integers but always use log(x, base = 2).

sjob <- slurm_apply(log,
                    data.frame(x = 1:10000),
                    base = 2,
                    nodes = 2, cpus_per_node = 2)

To pass additional objects to the jobs that aren't explicitly included as arguments to the function passed to slurm_apply or slurm_map, we can use the argument global_objects. For example we might want to use an inline function that calls two other previously defined functions.

sjob <- slurm_apply(function(a, b) c(func1(a),func2(b)), 
                    data.frame(a, b),
                    global_objects = c("func1", "func2"),
                    nodes = 2, cpus_per_node = 2)

The global_objects argument specifies the names of any R objects (besides the parameters data frame) that must be accessed by the function passed to slurm_apply. These objects are saved to a .RDS file that is loaded on each cluster node prior to evaluating the function in parallel.

By default, all R packages attached to the current R session will also be attached (with library) on each cluster node, though this can be modified with the optional pkgs argument.

Configuring Slurm options

Particular clusters may require the specification of additional Slurm options, such as time and memory limits for the job. The slurm_options argument allows you to set any of the command line options (view list) recognized by the Slurm sbatch command. It should be formatted as a named list, using the long names of each option (e.g. "time" rather than "t"). Flags, i.e. command line options that are toggled rather than set to a particular value, should be set to TRUE in slurm_options. For example, the following code sets the command line options --time=1:00:00 --share.

sopt <- list(time = '1:00:00', share = TRUE)
sjob <- slurm_apply(test_func, pars, slurm_options = sopt)

How it works / advanced customization

As mentioned above, the slurm_apply function creates a job-specific folder. This folder contains the parameters as a RDS file and (if applicable) the objects specified as global_objects saved together in a RData file. The function also generates a R script (slurm_run.R) to be run on each cluster node, as well as a Bash script ( to submit the job to Slurm.

More specifically, the Bash script tells Slurm to create a job array and the R script takes advantage of the unique SLURM\_ARRAY\_TASK\_ID environment variable that Slurm will set on each cluster node. This variable is read by slurm_run.R, which allows each instance of the script to operate on a different parameter subset and write its output to a different results file. The R script calls parallel::mcMap to parallelize calculations on each node.

Additionally, the --dependency option can be utilized by taking the job ID from the slurm_job object returned by slurm_apply, slurm_map, and slurm_call functions.
The ID can be manually added to the slurm options. In the following example, the job ID of sjob1 is used to ensure that sjob2 does not begin running until after sjob1 finishes.

# Job1
sopt1 <- list(time = '1:00:00', share = TRUE)
sjob1 <- slurm_apply(test_func, pars, slurm_options = sopt1)

# Job2 depends on Job1
pars2 <- data.frame(par_mu = 2:20,
                   par_sd = seq(0.2, 2, length.out = 20))
sopt2 <- c(sopt1, list(dependency=sprintf("afterany:%s", sjob1$jobid)))
sjob2 <- slurm_apply(test_func2, pars2, slurm_options = sopt2)

Both slurm_run.R and are generated from templates, using the whisker package; these templates can be found in the rslurm/templates subfolder in your R package library. There are two templates for each script, one for slurm_apply and the other (with the word "single"" in its title) for slurm_call.

While you should avoid changing any existing lines in the template scripts, you may want to add #SBATCH lines to the templates in order to permanently set certain Slurm command line options and thus customize the package to your particular cluster setup.