SnowParam-class: Enable simple network of workstations (SNOW)-style parallel...

In BiocParallel: Bioconductor facilities for parallel evaluation

Description

This class is used to parameterize simple network of workstations (SNOW) parallel evaluation on one or several physical computers. snowWorkers() chooses the number of workers.

Usage

## constructor
## ------------------------------------

SnowParam(workers = snowWorkers(type), type=c("SOCK", "MPI", "FORK"),
          tasks = 0L, stop.on.error = TRUE,
          progressbar = FALSE, RNGseed = NULL,
          timeout = 30L * 24L * 60L * 60L, exportglobals = TRUE,
          log = FALSE, threshold = "INFO", logdir = NA_character_,
          resultdir = NA_character_, jobname = "BPJOB",
          manager.hostname = NA_character_, manager.port = NA_integer_,
          ...)

## coercion
## ------------------------------------

## as(SOCKcluster, SnowParam)
## as(spawnedMPIcluster,SnowParam)

## detect workers
## ------------------------------------

snowWorkers(type = c("SOCK", "MPI", "FORK"))

Arguments

workers	integer(1) Number of workers. Defaults to all cores available as determined by detectCores. For a SOCK cluster workers can be a character() vector of host names.
type	character(1) Type of cluster to use. Possible values are SOCK (default) and MPI. Instead of type=FORK use MulticoreParam.
tasks	integer(1). The number of tasks per job. value must be a scalar integer >= 0L. In this documentation a job is defined as a single call to a function, such as bplapply, bpmapply etc. A task is the division of the X argument into chunks. When tasks == 0 (default), X is divided as evenly as possible over the number of workers. A tasks value of > 0 specifies the exact number of tasks. Values can range from 1 (all of X to a single worker) to the length of X (each element of X to a different worker). When the length of X is less than the number of workers each element of X is sent to a worker and tasks is ignored.
stop.on.error	logical(1) Enable stop on error.
progressbar	logical(1) Enable progress bar (based on plyr:::progress_text).
RNGseed	integer(1) Seed for random number generation. When not NULL, this value is passed to parallel::clusterSetRNGStream to generate random number streams on each worker.
timeout	numeric(1) Time (in seconds) allowed for worker to complete a task. This value is passed to base::setTimeLimit() as both the cpu and elapsed arguments. If the computation exceeds timeout an error is thrown with message 'reached elapsed time limit'.
exportglobals	logical(1) Export base::options() from manager to workers? Default TRUE.
log	logical(1) Enable logging.
threshold	character(1) Logging threshold as defined in futile.logger.
logdir	character(1) Log files directory. When not provided, log messages are returned to stdout.
resultdir	character(1) Job results directory. When not provided, results are returned as an R object (list) to the workspace.
jobname	character(1) Job name that is prepended to log and result files. Default is "BPJOB".
manager.hostname	character(1) Host name of manager node. See 'Global Options', below.
manager.port	integer(1) Port on manager with which workers communicate. See 'Global Options', below.
...	Additional arguments passed to makeCluster

Details

SnowParam is used for distributed memory computing and supports 2 cluster types: ‘SOCK’ (default) and ‘MPI’. The SnowParam builds on infrastructure in the snow and parallel packages and provides the additional features of error handling, logging and writing out results.

The default number of workers is determined by snowWorkers() which is usually the maximum of 1L and parallel::detectCores() - 2. Machines with 3 or fewer cores, or machines where number of cores cannot be determined, are assigned a single worker. Machines with more than 127 cores are limited to the number of R connections available when the workers start; this is 128 (a hard-coded limit in R) minus the number of open connections as returned by nrow(showConnections(all=TRUE)). The option mc.cores can be used to specify an arbitrary number of workers, e.g., options(mc.cores=4L); the Bioconductor build system enforces a maximum of 4 workers.

error handling:

By default all computations are attempted and partial results are returned with any error messages.

• stop.on.error A logical. Stops all jobs as soon as one job fails or wait for all jobs to terminate. When FALSE, the return value is a list of successful results along with error messages as 'conditions'.

• The bpok(x) function returns a logical() vector that is FALSE for any jobs that threw an error. The input x is a list output from a bp*apply function such as bplapply or bpmapply.

logging:

When log = TRUE the futile.logger package is loaded on the workers. All log messages written in the futile.logger format are captured by the logging mechanism and returned real-time (i.e., as each task completes) instead of after all jobs have finished.

Messages sent to stdout and stderr are returned to the workspace by default. When log = TRUE these are diverted to the log output. Those familiar with the outfile argument to makeCluster can think of log = FALSE as equivalent to outfile = NULL; providing a logdir is the same as providing a name for outfile except that BiocParallel writes a log file for each task.

The log output includes additional statistics such as memory use and task runtime. Memory use is computed by calling gc(reset=TRUE) before code evaluation and gc() (no reseet) after. The output of the second gc() call is sent to the log file. There are many ways to track memory use - this particular approach was taken because it is consistent with how the BatchJobs package reports memory on the workers.

log and result files:

Results and logs can be written to a file instead of returned to the workspace. Writing to files is done from the master as each task completes. Options can be set with the logdir and resultdir fields in the constructor or with the accessors, bplogdir and bpresultdir.

random number generation:

MulticoreParam and SnowParam use the random number generation support from the parallel package. These params are snow-derived clusters so the arguments for multicore-derived functions such as mc.set.seed and mc.reset.stream do not apply.

Random number generation is controlled through the param argument, RNGseed which is passed to parallel::clusterSetRNGStream. clusterSetRNGStream uses the L'Ecuyer-CMRG random number generator and distributes streams to the members of a cluster. If RNGseed is not NULL it serves as the seed to the streams, otherwise the streams are set from the current seed of the master process after selecting the L'Ecuyer generator. See ?clusterSetRNGStream for more details.

NOTE: The PSOCK cluster from the parallel package does not support cluster options scriptdir and useRscript. PSOCK is not supported because these options are needed to re-direct to an alternate worker script located in BiocParallel.

Constructor

SnowParam(workers = snowWorkers(), type=c("SOCK", "MPI"), tasks = 0L, stop.on.error = FALSE, progressbar = FALSE, RNGseed = NULL, timeout = Inf, exportglobals = TRUE, log = FALSE, threshold = "INFO", logdir = NA_character_, resultdir = NA_character_, jobname = "BPJOB", manager.hostname = NA_character_, manager.port = NA_integer_, ...):

Return an object representing a SNOW cluster. The cluster is not created until bpstart is called. Named arguments in ... are passed to makeCluster.

Accessors: Logging and results

In the following code, x is a SnowParam object.

bpprogressbar(x), bpprogressbar(x) <- value: Get or set the value to enable text progress bar. value must be a logical(1).

bpjobname(x), bpjobname(x) <- value: Get or set the job name.

bpRNGseed(x), bpRNGseed(x) <- value: Get or set the seed for random number generaton. value must be a numeric(1) or NULL.

bplog(x), bplog(x) <- value: Get or set the value to enable logging. value must be a logical(1).

bpthreshold(x), bpthreshold(x) <- value: Get or set the logging threshold. value must be a character(1) string of one of the levels defined in the futile.logger package: “TRACE”, “DEBUG”, “INFO”, “WARN”, “ERROR”, or “FATAL”.

bplogdir(x), bplogdir(x) <- value: Get or set the directory for the log file. value must be a character(1) path, not a file name. The file is written out as BPLOG.out. If no logdir is provided and bplog=TRUE log messages are sent to stdout.

bpresultdir(x), bpresultdir(x) <- value: Get or set the directory for the result files. value must be a character(1) path, not a file name. Separate files are written for each job with the prefix TASK (e.g., TASK1, TASK2, etc.). When no resultdir is provided the results are returned to the session as list.

Accessors: Back-end control

In the code below x is a SnowParam object. See the ?BiocParallelParam man page for details on these accessors.

bpworkers(x), bpworkers(x) <- value, bpnworkers(x)

bptasks(x), bptasks(x) <- value

bpstart(x)

bpstop(x)

bpisup(x)

bpbackend(x), bpbackend(x) <- value

Accessors: Error Handling

In the code below x is a SnowParam object. See the ?BiocParallelParam man page for details on these accessors.

bpstopOnError(x), bpstopOnError(x) <- value

Methods: Evaluation

In the code below BPPARAM is a SnowParam object. Full documentation for these functions are on separate man pages: see ?bpmapply, ?bplapply, ?bpvec, ?bpiterate and ?bpaggregate.

bpmapply(FUN, ..., MoreArgs=NULL, SIMPLIFY=TRUE, USE.NAMES=TRUE, BPPARAM=bpparam())

bplapply(X, FUN, ..., BPPARAM=bpparam())

bpvec(X, FUN, ..., AGGREGATE=c, BPPARAM=bpparam())

bpiterate(ITER, FUN, ..., BPPARAM=bpparam())

bpaggregate(x, data, FUN, ..., BPPARAM=bpparam())

Methods: Other

In the code below x is a SnowParam object.

show(x): Displays the SnowParam object.

bpok(x): Returns a logical() vector: FALSE for any jobs that resulted in an error. x is the result list output by a BiocParallel function such as bplapply or bpmapply.

Coercion

as(from, "SnowParam"): Creates a SnowParam object from a SOCKcluster or spawnedMPIcluster object. Instances created in this way cannot be started or stopped.

Global Options

The global option mc.cores influences the number of workers determined by snowWorkers() (described above) or multicoreWorkers() (see multicoreWorkers).

Workers communicate to the master through socket connections. Socket connections require a hostname and port. These are determined by arguments manager.hostname and manager.port; default values are influenced by global options.

The default manager hostname is "localhost" when the number of workers are specified as a numeric(1), and Sys.info()[["nodename"]] otherwise. The hostname can be over-ridden by the envirnoment variable MASTER, or the global option bphost (e.g., options(bphost=Sys.info()[["nodename"]]).

The default port is chosen as a random value between 11000 and 11999. The port may be over-ridden by the environment variable R_PARALLEL_PORT or PORT, and by the option ports, e.g., options(ports=12345L).

Author(s)

Martin Morgan and Valerie Obenchain.

See Also

• register for registering parameter classes for use in parallel evaluation.

• MulticoreParam for computing in shared memory

• BatchJobsParam for computing with cluster schedulers

• DoparParam for computing with foreach

• SerialParam for non-parallel evaluation

Examples

## -----------------------------------------------------------------------
## Job configuration:
## -----------------------------------------------------------------------

## SnowParam supports distributed memory computing. The object fields
## control the division of tasks, error handling, logging and result
## format.
bpparam <- SnowParam()
bpparam

## Fields are modified with accessors of the same name:
bplog(bpparam) <- TRUE
dir.create(resultdir <- tempfile())
bpresultdir(bpparam) <- resultdir
bpparam

## -----------------------------------------------------------------------
## Logging:
## -----------------------------------------------------------------------

## When 'log == TRUE' the workers use a custom script (in BiocParallel)
## that enables logging and access to other job statistics. Log messages
## are returned as each job completes rather than waiting for all to
## finish.

## In 'fun', a value of 'x = 1' will throw a warning, 'x = 2' is ok
## and 'x = 3' throws an error. Because 'x = 1' sleeps, the warning
## should return after the error.

X <- 1:3
fun <- function(x) {
    if (x == 1) {
        Sys.sleep(2)
        if (TRUE & c(TRUE, TRUE))  ## warning
            x
    } else if (x == 2) {
        x                          ## ok
    } else if (x == 3) {
        sqrt("FOO")                ## error
    }
}

## By default logging is off. Turn it on with the bplog()<- setter
## or by specifying 'log = TRUE' in the constructor.
bpparam <- SnowParam(3, log = TRUE, stop.on.error = FALSE)
tryCatch({
    bplapply(X, fun, BPPARAM = bpparam)
}, error=identity)

## When a 'logdir' location is given the messages are redirected to a
## file:
## Not run: 
dir.create(logdir <- tempfile())
bplogdir(bpparam) <- logdir
bplapply(X, fun, BPPARAM = bpparam)
list.files(bplogdir(bpparam))

## End(Not run)

## -----------------------------------------------------------------------
## Managing results:
## -----------------------------------------------------------------------

## By default results are returned as a list. When 'resultdir' is given
## files are saved in the directory specified by job, e.g., 'TASK1.Rda',
## 'TASK2.Rda', etc.
## Not run: 
dir.create(resultdir <- tempfile())
bpparam <- SnowParam(2, resultdir = resultdir)
bplapply(X, fun, BPPARAM = bpparam)
list.files(bpresultdir(bpparam))

## End(Not run)

## -----------------------------------------------------------------------
## Error handling:
## -----------------------------------------------------------------------

## When 'stop.on.error' is TRUE the process returns as soon as an error
## is thrown.

## When 'stop.on.error' is FALSE all computations are attempted. Partial
## results are returned along with errors. Use bptry() to see the
## partial results
bpparam <- SnowParam(2, stop.on.error = FALSE)
res <- bptry(bplapply(list(1, "two", 3, 4), sqrt, BPPARAM = bpparam))
res

## Calling bpok() on the result list returns TRUE for elements with no
## error.
bpok(res)

## -----------------------------------------------------------------------
## Random number generation:
## -----------------------------------------------------------------------

## Random number generation is controlled with the 'RNGseed' field.
## This seed is passed to parallel::clusterSetRNGStream
## which uses the L'Ecuyer-CMRG random number generator and distributes
## streams to members of the cluster.

bpparam <- SnowParam(3, RNGseed = 7739465)
bplapply(seq_len(bpnworkers(bpparam)), function(i) rnorm(1),
         BPPARAM = bpparam)

BiocParallel documentation built on Nov. 8, 2020, 5:46 p.m.