Use slurm_apply
to compute function over multiple sets of
parameters in parallel, spread across multiple nodes of a Slurm cluster,
with similar syntax to mapply
.
slurm_apply(
f,
params,
...,
jobname = NA,
nodes = 2,
cpus_per_node = 2,
processes_per_node = cpus_per_node,
preschedule_cores = TRUE,
job_array_task_limit = NULL,
global_objects = NULL,
add_objects = NULL,
pkgs = rev(.packages()),
libPaths = NULL,
rscript_path = NULL,
r_template = NULL,
sh_template = NULL,
slurm_options = list(),
submit = TRUE
)
A function that accepts one or many single values as parameters and may return any type of R object.
A data frame of parameter values to apply f
to. Each
column corresponds to a parameter of f
(Note: names must
match) and each row corresponds to a separate function call.
Additional arguments to f
. These arguments do not vary
with each call to f
.
The name of the Slurm job; if NA
, it is assigned a
random name of the form "slr####".
The (maximum) number of cluster nodes to spread the calculation
over. slurm_apply
automatically divides params
in chunks of
approximately equal size to send to each node. Less nodes are allocated if
the parameter set is too small to use all CPUs on the requested nodes.
The number of CPUs requested per node. This argument is
mapped to the Slurm parameter cpus-per-task
.
The number of logical CPUs to utilize per node,
i.e. how many processes to run in parallel per node. This can exceed
cpus_per_node
for nodes which support hyperthreading. Defaults to
processes_per_node = cpus_per_node
.
Corresponds to the mc.preschedule
argument of
parallel::mcmapply
. Defaults to TRUE
. If TRUE
, the
rows of params
are assigned to cores before computation. If FALSE
,
each row of params
is executed by the next available core.
Setting FALSE
may be faster if
different values of params
result in very variable completion time for
jobs.
The maximum number of job array tasks to run at
the same time. Defaults to NULL
(no limit).
A character vector containing the name of R objects to be
saved in a .RData file and loaded on each cluster node prior to calling
f
.
Older deprecated name of global_objects
, retained for
backwards compatibility.
A character vector containing the names of packages that must
be loaded on each cluster node. By default, it includes all packages
loaded by the user when slurm_apply
is called.
A character vector describing the location of additional R
library trees to search through, or NULL. The default value of NULL
corresponds to libraries returned by .libPaths()
on a cluster node.
Non-existent library trees are silently ignored.
The location of the Rscript command. If not specified, defaults to the location of Rscript within the R installation being run.
The path to the template file for the R script run on each node. If NULL, uses the default template "rslurm/templates/slurm_run_R.txt".
The path to the template file for the sbatch submission script. If NULL, uses the default template "rslurm/templates/submit_sh.txt".
A named list of options recognized by sbatch
; see
Details below for more information.
Whether or not to submit the job to the cluster with
sbatch
; see Details below for more information.
A slurm_job
object containing the jobname
and the
number of nodes
effectively used.
This function creates a temporary folder ("_rslurm_[jobname]") in the current directory, holding .RData and .RDS data files, the R script to run and the Bash submission script generated for the Slurm job.
The set of input parameters is divided in equal chunks sent to each node, and
f
is evaluated in parallel within each node using functions from the
parallel
R package. The names of any other R objects (besides
params
) that f
needs to access should be included in
global_objects
or passed as additional arguments through ...
.
Use slurm_options
to set any option recognized by sbatch
, e.g.
slurm_options = list(time = "1:00:00", share = TRUE)
.
See http://slurm.schedmd.com/sbatch.html for details on possible options.
Note that full names must be used (e.g. "time" rather than "t") and that flags
(such as "share") must be specified as TRUE. The "array", "job-name", "nodes",
"cpus-per-task" and "output" options are already determined by
slurm_apply
and should not be manually set.
When processing the computation job, the Slurm cluster will output two types of files in the temporary folder: those containing the return values of the function for each subset of parameters ("results_[node_id].RDS") and those containing any console or error output produced by R on each node ("slurm_[node_id].out").
If submit = TRUE
, the job is sent to the cluster and a confirmation
message (or error) is output to the console. If submit = FALSE
,
a message indicates the location of the saved data and script files; the
job can be submitted manually by running the shell command
sbatch submit.sh
from that directory.
After sending the job to the Slurm cluster, slurm_apply
returns a
slurm_job
object which can be used to cancel the job, get the job
status or output, and delete the temporary files associated with it. See
the description of the related functions for more details.
slurm_call
to evaluate a single function call.
slurm_map
to evaluate a function over a list.
cancel_slurm
, cleanup_files
,
get_slurm_out
and get_job_status
which use the output of this function.
if (FALSE) {
sjob <- slurm_apply(func, pars)
get_job_status(sjob) # Prints console/error output once job is completed.
func_result <- get_slurm_out(sjob, "table") # Loads output data into R.
cleanup_files(sjob)
}