parallelMap: parallelStart – R documentation

Become an expert in R — Interactive courses, Cheat Sheets, certificates and more!

parallelStart

Parallelization setup for parallelMap.

Description

Defines the underlying parallelization mode for parallelMap(). Also allows to set a “level” of parallelization. Only calls to parallelMap() with a matching level are parallelized. The defaults of all settings are taken from your options, which you can also define in your R profile. For an introductory tutorial and information on the options configuration, please go to the project's github page at https://github.com/mlr-org/parallelMap.

Usage

parallelStart(
  mode,
  cpus,
  socket.hosts,
  bj.resources = list(),
  bt.resources = list(),
  logging,
  storagedir,
  level,
  load.balancing = FALSE,
  show.info,
  suppress.local.errors = FALSE,
  reproducible,
  ...
)

parallelStartLocal(show.info, suppress.local.errors = FALSE, ...)

parallelStartMulticore(
  cpus,
  logging,
  storagedir,
  level,
  load.balancing = FALSE,
  show.info,
  reproducible,
  ...
)

parallelStartSocket(
  cpus,
  socket.hosts,
  logging,
  storagedir,
  level,
  load.balancing = FALSE,
  show.info,
  reproducible,
  ...
)

parallelStartMPI(
  cpus,
  logging,
  storagedir,
  level,
  load.balancing = FALSE,
  show.info,
  reproducible,
  ...
)

parallelStartBatchJobs(
  bj.resources = list(),
  logging,
  storagedir,
  level,
  show.info,
  ...
)

parallelStartBatchtools(
  bt.resources = list(),
  logging,
  storagedir,
  level,
  show.info,
  ...
)

Arguments

`mode`	(`character(1)`) Which parallel mode should be used: “local”, “multicore”, “socket”, “mpi”, “BatchJobs”. Default is the option `parallelMap.default.mode` or, if not set, “local” without parallel execution.
`cpus`	(`integer(1)`) Number of used cpus. For local and BatchJobs mode this argument is ignored. For socket mode, this is the number of processes spawned on localhost, if you want processes on multiple machines use `socket.hosts`. Default is the option `parallelMap.default.cpus` or, if not set, `parallel::detectCores()` for multicore mode, `max(1, [mpi.universe.size][Rmpi::mpi.universe.size] - 1)` for mpi mode and 1 for socket mode.
`socket.hosts`	character Only used in socket mode, otherwise ignored. Names of hosts where parallel processes are spawned. Default is the option `parallelMap.default.socket.hosts`, if this option exists.
`bj.resources`	list Resources like walltime for submitting jobs on HPC clusters via BatchJobs. See `BatchJobs::submitJobs()`. Defaults are taken from your BatchJobs config file.
`bt.resources`	list Analog to `bj.resources`. See `batchtools::submitJobs()`.
`logging`	(`logical(1)`) Should slave output be logged to files via `sink()` under the `storagedir`? Files are named `<iteration_number>.log` and put into unique subdirectories named `parallelMap_log_<nr>` for each subsequent `parallelMap()` operation. Previous logging directories are removed on `parallelStart` if `logging` is enabled. Logging is not supported for local mode, because you will see all output on the master and can also run stuff like `traceback()` in case of errors. Default is the option `parallelMap.default.logging` or, if not set, `FALSE`.
`storagedir`	(`character(1)`) Existing directory where log files and intermediate objects for BatchJobs mode are stored. Note that all nodes must have write access to exactly this path. Default is the current working directory.
`level`	(`character(1)`) You can set this so only calls to `parallelMap()` that have exactly the same level are parallelized. Default is the option `parallelMap.default.level` or, if not set, `NA` which means all calls to `parallelMap()` are are potentially parallelized.
`load.balancing`	(`logical(1)`) Enables load balancing for multicore, socket and mpi. Set this to `TRUE` if you have heterogeneous runtimes. Default is `FALSE`
`show.info`	(`logical(1)`) Verbose output on console for all further package calls? Default is the option `parallelMap.default.show.info` or, if not set, `TRUE`.
`suppress.local.errors`	(`logical(1)`) Should reporting of error messages during function evaluations in local mode be suppressed? Default ist FALSE, i.e. every error message is shown.
`reproducible`	(`logical(1)`) Should parallel jobs produce reproducible results when setting a seed? With this option, `parallelMap()` calls will be reproducible when using `set.seed()` with the default RNG kind. This is not the case by default when parallelizing in R, since the default RNG kind "Mersenne-Twister" is not honored by parallel processes. Instead RNG kind `"L'Ecuyer-CMRG"` needs to be used to ensure paralllel reproducibility. Default is the option `parallelMap.default.reproducible` or, if not set, `TRUE`.
`...`	(any) Optional parameters, for socket mode passed to `parallel::makePSOCKcluster()`, for mpi mode passed to `parallel::makeCluster()` and for multicore passed to `parallel::mcmapply()` (`mc.preschedule` (overwriting `load.balancing`), `mc.set.seed`, `mc.silent` and `mc.cleanup` are supported for multicore).

Details

Currently the following modes are supported, which internally dispatch the mapping operation to functions from different parallelization packages:

local: No parallelization with mapply().
multicore: Multicore execution on a single machine with parallel::mclapply().
mpi: Snow MPI cluster on one or multiple machines with parallel::makeCluster() and parallel::clusterMap().
BatchJobs: Parallelization on batch queuing HPC clusters, e.g., Torque, SLURM, etc., with BatchJobs::batchMap().

For BatchJobs mode you need to define a storage directory through the argument storagedir or the option parallelMap.default.storagedir.

Value

Nothing.

parallelMap

Unified Interface to Parallelization Back-Ends

v1.5.0

BSD_2_clause + file LICENSE

Authors

Bernd Bischl [cre, aut], Michel Lang [aut] (<https://orcid.org/0000-0001-9754-0393>), Patrick Schratz [aut] (<https://orcid.org/0000-0003-0748-6624>)

Initial release