Parallelize a Vector Map Function using Forking

Description

pvec parallelizes the execution of a function on vector elements by splitting the vector and submitting each part to one core. The function must be a vectorized map, i.e. it takes a vector input and creates a vector output of exactly the same length as the input which doesn't depend on the partition of the vector.

It relies on forking and hence is not available on Windows unless mc.cores = 1.

Usage

pvec(v, FUN, ..., mc.set.seed = TRUE, mc.silent = FALSE,
     mc.cores = getOption("mc.cores", 2L), mc.cleanup = TRUE)

Arguments

Details

pvec parallelizes FUN(x, ...) where FUN is a function that returns a vector of the same length as x. FUN must also be pure (i.e., without side-effects) since side-effects are not collected from the parallel processes. The vector is split into nearly identically sized subvectors on which FUN is run. Although it is in principle possible to use functions that are not necessarily maps, the interpretation would be case-specific as the splitting is in theory arbitrary (a warning is given in such cases).

The major difference between pvec and mclapply is that mclapply will run FUN on each element separately whereas pvec assumes that c(FUN(x[1]), FUN(x[2])) is equivalent to FUN(x[1:2]) and thus will split into as many calls to FUN as there are cores (or elements, if fewer), each handling a subset vector. This makes it more efficient than mclapply but requires the above assumption on FUN.

If mc.cores == 1 this evaluates FUN(v, ...) in the current process.

Value

The result of the computation – in a successful case it should be of the same length as v. If an error occurred or the function was not a map the result may be shorter or longer, and a warning is given.

Note

Due to the nature of the parallelization, error handling does not follow the usual rules since errors will be returned as strings and results from killed child processes will show up simply as non-existent data. Therefore it is the responsibility of the user to check the length of the result to make sure it is of the correct size. pvec raises a warning if that is the case since it does not know whether such an outcome is intentional or not.

See mcfork for the inadvisability of using this with GUI front-ends and multi-threaded libraries.

Author(s)

Simon Urbanek and R Core.

Derived from the multicore package formerly on CRAN.

See Also

mcparallel, mclapply, parLapply, clusterMap.

Examples

x <- pvec(1:1000, sqrt)
stopifnot(all(x == sqrt(1:1000)))


# One use is to convert date strings to unix time in large datasets
# as that is a relatively slow operation.
# So let's get some random dates first
# (A small test only with 2 cores: set options("mc.cores")
# and increase N for a larger-scale test.)
N <- 1e5
dates <- sprintf('%04d-%02d-%02d', as.integer(2000+rnorm(N)),
                 as.integer(runif(N, 1, 12)), as.integer(runif(N, 1, 28)))

system.time(a <- as.POSIXct(dates))

# But specifying the format is faster
system.time(a <- as.POSIXct(dates, format = "%Y-%m-%d"))

# pvec ought to be faster, but system overhead can be high
system.time(b <- pvec(dates, as.POSIXct, format = "%Y-%m-%d"))
stopifnot(all(a == b))

# using mclapply for this would much slower because each value
# will require a separate call to as.POSIXct()
# as lapply(dates, as.POSIXct) does
system.time(c <- unlist(mclapply(dates, as.POSIXct,  format = "%Y-%m-%d")))
stopifnot(all(a == c))