The future package provides a generic API for using futures in R. A future is a simple yet powerful mechanism to evaluate an R expression and retrieve its value at some point in time. Futures can be resolved in many different ways depending on which strategy is used. There are various types of synchronous and asynchronous futures to choose from in the future package. Additional futures are implemented in other packages. For instance, the future.batchtools package provides futures for any type of backend that the batchtools package supports. For an introduction to futures in R, please consult the vignettes of the future package.
The doFuture package provides a %dopar% adapter for the foreach
package that works with any type of future.
The doFuture package is cross platform just as the future package.
Below is an example showing how to make %dopar% work with
multiprocess futures. A multiprocess future will be evaluated in
parallel using forked processes. If process forking is not supported
by the operating system, then multiple background R sessions will
instead be used to resolve the futures.
library("doFuture")
registerDoFuture()
plan(multiprocess)
mu <- 1.0
sigma <- 2.0
x <- foreach(i = 1:3, .export = c("mu", "sigma")) %dopar% {
rnorm(i, mean = mu, sd = sigma)
}
To do the same on high-performance computing (HPC) cluster, the future.batchtools package can be used. Assuming batchtools has been configured correctly, then following foreach iterations will be submitted to the HPC job scheduler and distributed for evaluation on the compute nodes.
library("doFuture")
registerDoFuture()
library("future.batchtools")
plan(batchjobs_slurm)
mu <- 1.0
sigma <- 2.0
x <- foreach(i = 1:3, .export = c("mu", "sigma")) %dopar% {
rnorm(i, mean = mu, sd = sigma)
}
The plyr package uses foreach as a parallel backend. This means that with doFuture any type of futures can be used for asynchronous (and synchronous) plyr processing including multicore, multisession, MPI, ad hoc clusters and HPC job schedulers. For example,
library("doFuture")
registerDoFuture()
plan(multiprocess)
library("plyr")
x <- list(a = 1:10, beta = exp(-3:3), logic = c(TRUE, FALSE, FALSE, TRUE))
y <- llply(x, quantile, probs = (1:3) / 4, .parallel = TRUE)
## $a
## 25% 50% 75%
## 3.25 5.50 7.75
##
## $beta
## 25% 50% 75%
## 0.2516074 1.0000000 5.0536690
##
## $logic
## 25% 50% 75%
## 0.0 0.5 1.0
The BiocParallel package supports any %dopar% adapter as a parallel backend. This means that with doFuture, BiocParallel supports any type of future. For example,
library("doFuture")
registerDoFuture()
plan(multiprocess)
library("BiocParallel")
register(DoparParam(), default = TRUE)
mu <- 1.0
sigma <- 2.0
x <- bplapply(1:3, mu = mu, sigma = sigma, function(i, mu, sigma) {
rnorm(i, mean = mu, sd = sigma)
})
The foreach package itself has some support for automated handling of globals but unfortunately it does not work in all cases. Specifically, if foreach() is called from within a function, you do need to export globals explicitly. For example, although globals my and sigma are properly exported when we do
> library("doParallel")
> registerDoParallel(parallel::makeCluster(2))
> mu <- 1.0
> sigma <- 2.0
> x <- foreach(i = 1:3) %dopar% { rnorm(i, mean = mu, sd = sigma) }
> str(x)
List of 3
$ : num -1.42
$ : num [1:2] 3.12 -1.33
$ : num [1:3] -0.0376 -0.1446 1.6368
it falls short as soon as we try to do the same from within a function;
> foo <- function() foreach(i = 1:3) %dopar% { rnorm(i, mean = mu, sd = sigma) }
> x <- foo()
Error in { : task 1 failed - "object 'mu' not found"
The solution is to explicitly export global variables, e.g.
> foo <- function() {
+ foreach(i = 1:3, .export = c("mu", "sigma")) %dopar% {
+ rnorm(i, mean = mu, sd = sigma)
+ }
+ }
> x <- foo()
However, when using the %dopar% adapter of doFuture, all of the future machinery comes in to play including automatic handling of global variables, e.g.
> library("doFuture")
> registerDoFuture()
> plan(multisession, workers = 2)
> mu <- 1.0
> sigma <- 2.0
> foo <- function() foreach(i = 1:3) %dopar% { rnorm(i, mean = mu, sd = sigma) }
> x <- foo()
> str(x)
List of 3
$ : num 0.358
$ : num [1:2] 3.317 -0.689
$ : num [1:3] -0.104 1.237 2.474
Another advantage with doFuture is that, contrary to doParallel, packages that need to be attached are also automatically taken care of, e.g.
> registerDoFuture()
> library("tools")
> ext <- foreach(file = c("abc.txt", "def.log")) %dopar% file_ext(file)
> unlist(ext)
[1] "txt" "log"
whereas
> registerDoParallel(parallel::makeCluster(2))
> library("tools")
> ext <- foreach(file = c("abc.txt", "def.log")) %dopar% file_ext(file)
Error in file_ext(file) :
task 1 failed - "could not find function "file_ext""
Having said all this, in order to write foreach code that works everywhere, it is better to be conservative and not assume that all end users will use a doFuture backend. Because of this, it is still recommended to explicitly specify all objects that need to be export whenever using the foreach API. The doFuture framework can help you identify what should go into the .export argument. By setting options(doFuture.foreach.export = ".export-and-automatic-with-warning"), doFuture will in warn if it finds globals not listed in .export and produce an informative warning message suggesting that those should be added. To assert that argument .export is correct, test the code with options(doFuture.foreach.export = ".export"), which will disable automatic identification of globals such that only the globals specified by the .export argument is used.
Due to the generic nature of futures, the doFuture package provides the same functionality as many of the existing doNnn packages combined, e.g. doMC, doParallel, doMPI, and doSNOW.
| doNnn usage | doFuture alternative |
|---|---|
|
|
|
|
|
|
| N/A |
|
|
|
|
|
|
|
| N/A |
High-performance compute (HPC) schedulers, e.g.
SGE, Slurm, and TORQUE / PBS.
|
|
N/A. There is currently no known Redis-based future backend and therefore no known doFuture alternative to the doRedis package. |