submitJobs {batchtools}R Documentation

Submit Jobs to the Batch Systems


Submits defined jobs to the batch system.

After submitting the jobs, you can use waitForJobs to wait for the termination of jobs or call reduceResultsList/reduceResults to collect partial results. The progress can be monitored with getStatus.


  ids = NULL,
  resources = list(),
  sleep = NULL,
  reg = getDefaultRegistry()



[data.frame or integer]
A data.frame (or data.table) with a column named “”. Alternatively, you may also pass a vector of integerish job ids. If not set, defaults to the return value of findNotSubmitted. Invalid ids are ignored.


[named list]
Computational resources for the jobs to submit. The actual elements of this list (e.g. something like “walltime” or “nodes”) depend on your template file, exceptions are outlined in the section 'Resources'. Default settings for a system can be set in the configuration file by defining the named list default.resources. Note that these settings are merged by name, e.g. merging list(walltime = 300) into list(walltime = 400, memory = 512) will result in list(walltime = 300, memory = 512). Same holds for individual job resources passed as additional column of ids (c.f. section 'Resources').


[function(i) | numeric(1)]
Parameter to control the duration to sleep between temporary errors. You can pass an absolute numeric value in seconds or a function(i) which returns the number of seconds to sleep in the i-th iteration between temporary errors. If not provided (NULL), tries to read the value (number/function) from the configuration file (stored in reg$sleep) or defaults to a function with exponential backoff between 5 and 120 seconds.


Registry. If not explicitly passed, uses the default registry (see setDefaultRegistry).


[data.table] with columns “” and “chunk”.


You can pass arbitrary resources to submitJobs() which then are available in the cluster function template. Some resources' names are standardized and it is good practice to stick to the following nomenclature to avoid confusion:


Upper time limit in seconds for jobs before they get killed by the scheduler. Can be passed as additional column as part of ids to set per-job resources.


Memory limit in Mb. If jobs exceed this limit, they are usually killed by the scheduler. Can be passed as additional column as part of ids to set per-job resources.


Number of (physical) CPUs to use on the slave. Can be passed as additional column as part of ids to set per-job resources.


Number of threads to use via OpenMP. Used to set environment variable “OMP_NUM_THREADS”. Can be passed as additional column as part of ids to set per-job resources.


Maximum size of the pointer protection stack, see Memory.


Number of threads to use for the BLAS backend. Used to set environment variables “MKL_NUM_THREADS” and “OPENBLAS_NUM_THREADS”. Can be passed as additional column as part of ids to set per-job resources.


Enable memory measurement for jobs. Comes with a small runtime overhead.

Execute chunks as array jobs.


Start a parallelMap backend on the slave.


Start a foreach backend on the slave.


Resource used for Slurm to select the set of clusters to run sbatch/squeue/scancel on.

Chunking of Jobs

Multiple jobs can be grouped (chunked) together to be executed sequentially on the batch system as a single batch job. This is especially useful to avoid overburding the scheduler by submitting thousands of jobs simultaneously. To chunk jobs together, job ids must be provided as data.frame with columns “” and “chunk” (integer). All jobs with the same chunk number will be executed sequentially inside the same batch job. The utility functions chunk, binpack and lpt can assist in grouping jobs.

Array Jobs

If your cluster supports array jobs, you can set the resource to TRUE in order to execute chunks as job arrays on the cluster. For each chunk of size n, batchtools creates a JobCollection of (possibly heterogeneous) jobs which is submitted to the scheduler as a single array job with n repetitions. For each repetition, the JobCollection is first read from the file system, then subsetted to the i-th job using the environment variable reg$cluster.functions$array.var (depending on the cluster backend, defined automatically) and finally executed.

Order of Submission

Jobs are submitted in the order of chunks, i.e. jobs which have chunk number sort(unique(ids$chunk))[1] first, then jobs with chunk number sort(unique(ids$chunk))[2] and so on. If no chunks are provided, jobs are submitted in the order of ids$

Limiting the Number of Jobs

If requested, submitJobs tries to limit the number of concurrent jobs of the user by waiting until jobs terminate before submitting new ones. This can be controlled by setting “” in the configuration file (see Registry) or by setting the resource “” to the maximum number of jobs to run simultaneously. If both are set, the setting via the resource takes precedence over the setting in the configuration.

Measuring Memory

Setting the resource measure.memory to TRUE turns on memory measurement: gc is called directly before and after the job and the difference is stored in the internal database. Note that this is just a rough estimate and does neither work reliably for external code like C/C++ nor in combination with threading.

Inner Parallelization

Inner parallelization is typically done via threading, sockets or MPI. Two backends are supported to assist in setting up inner parallelization.

The first package is parallelMap. If you set the resource “pm.backend” to “multicore”, “socket” or “mpi”, parallelStart is called on the slave before the first job in the chunk is started and parallelStop is called after the last job terminated. This way, the resources for inner parallelization can be set and get automatically stored just like other computational resources. The function provided by the user just has to call parallelMap to start parallelization using the preconfigured backend.

To control the number of CPUs, you have to set the resource ncpus. Otherwise ncpus defaults to the number of available CPUs (as reported by (see detectCores)) on the executing machine for multicore and socket mode and defaults to the return value of mpi.universe.size-1 for MPI. Your template must be set up to handle the parallelization, e.g. request the right number of CPUs or start R with mpirun. You may pass further options like level to parallelStart via the named list “pm.opts”.

The second supported parallelization backend is foreach. If you set the resource “foreach.backend” to “seq” (sequential mode), “parallel” (doParallel) or “mpi” (doMPI), the requested foreach backend is automatically registered on the slave. Again, the resource ncpus is used to determine the number of CPUs.

Neither the namespace of parallelMap nor the namespace foreach are attached. You have to do this manually via library or let the registry load the packages for you.


If you a large number of jobs, disabling the progress bar (options(batchtools.progress = FALSE)) can significantly increase the performance of submitJobs.


### Example 1: Submit subsets of jobs
tmp = makeRegistry(file.dir = NA, make.default = FALSE)

# toy function which fails if x is even and an input file does not exists
fun = function(x, fn) if (x %% 2 == 0 && !file.exists(fn)) stop("file not found") else x

# define jobs via batchMap
fn = tempfile()
ids = batchMap(fun, 1:20, reg = tmp, fn = fn)

# submit some jobs
ids = 1:10
submitJobs(ids, reg = tmp)
waitForJobs(ids, reg = tmp)
getStatus(reg = tmp)

# create the required file and re-submit failed jobs
submitJobs(findErrors(ids, reg = tmp), reg = tmp)
getStatus(reg = tmp)

# submit remaining jobs which have not yet been submitted
ids = findNotSubmitted(reg = tmp)
submitJobs(ids, reg = tmp)
getStatus(reg = tmp)

# collect results
reduceResultsList(reg = tmp)

### Example 2: Using memory measurement
tmp = makeRegistry(file.dir = NA, make.default = FALSE)

# Toy function which creates a large matrix and returns the column sums
fun = function(n, p) colMeans(matrix(runif(n*p), n, p))

# Arguments to fun:
args = data.table::CJ(n = c(1e4, 1e5), p = c(10, 50)) # like expand.grid()

# Map function to create jobs
ids = batchMap(fun, args = args, reg = tmp)

# Set resources: enable memory measurement
res = list(measure.memory = TRUE)

# Submit jobs using the currently configured cluster functions
submitJobs(ids, resources = res, reg = tmp)

# Retrive information about memory, combine with parameters
info = ijoin(getJobStatus(reg = tmp)[, .(, mem.used)], getJobPars(reg = tmp))

# Combine job info with results -> each job is aggregated using mean()
unwrap(ijoin(info, reduceResultsDataTable(fun = function(res) list(res = mean(res)), reg = tmp)))

### Example 3: Multicore execution on the slave
tmp = makeRegistry(file.dir = NA, make.default = FALSE)

# Function which sleeps 10 seconds, i-times
f = function(i) {
  parallelMap::parallelMap(Sys.sleep, rep(10, i))

# Create one job with parameter i=4
ids = batchMap(f, i = 4, reg = tmp)

# Set resources: Use parallelMap in multicore mode with 4 CPUs
# batchtools internally loads the namespace of parallelMap and then
# calls parallelStart() before the job and parallelStop() right
# after the job last job in the chunk terminated.
res = list(pm.backend = "multicore", ncpus = 4)

## Not run: 
# Submit both jobs and wait for them
submitJobs(resources = res, reg = tmp)
waitForJobs(reg = tmp)

# If successfull, the running time should be ~10s
getJobTable(reg = tmp)[, .(, time.running)]

# There should also be a note in the log:
grepLogs(pattern = "parallelMap", reg = tmp)

## End(Not run)

[Package batchtools version 0.9.17 Index]