The goal of furrr is to combine purrr’s family of mapping functions with future’s parallel processing capabilities. The result is near drop in replacements for purrr functions such as
map2_dbl(), which can be replaced with their furrr equivalents of
future_map2_dbl() to map in parallel.
The code draws heavily from the implementations of purrr and future.apply and this package would not be possible without either of them.
Every variant of the following functions has been implemented:
You can install the released version of furrr from CRAN with:
And the development version from GitHub with:
# install.packages("remotes") remotes::install_github("DavisVaughan/furrr")
The easiest way to learn about furrr is to browse the website. In particular, the function reference page can be useful to get a general overview of the functions in the package, and the following vignettes are deep dives into various parts of furrr:
furrr has been designed to function as identically to purrr as possible, so that you can immediately have familiarity with it.
The default backend for future (and through it, furrr) is a sequential one. This means that the above code will run out of the box, but it will not be in parallel. The design of future makes it incredibly easy to change this so that your code will run in parallel.
# Set a "plan" for how the code should run. plan(multisession, workers = 2) # This does run in parallel! future_map(c("hello", "world"), ~.x) #> [] #>  "hello" #> #> [] #>  "world"
If you are still skeptical, here is some proof that we are running in parallel.
library(tictoc) # This should take 6 seconds in total running sequentially plan(sequential) tic() nothingness <- future_map(c(2, 2, 2), ~Sys.sleep(.x)) toc() #> 6.08 sec elapsed
# This should take ~2 seconds running in parallel, with a little overhead # in `future_map()` from sending data to the workers. There is generally also # a one time cost from `plan(multisession)` setting up the workers. plan(multisession, workers = 3) tic() nothingness <- future_map(c(2, 2, 2), ~Sys.sleep(.x)) toc() #> 2.212 sec elapsed
It’s important to remember that data has to be passed back and forth between the workers. This means that whatever performance gain you might have gotten from your parallelization can be crushed by moving large amounts of data around. For example, if you are moving large data frames to the workers, running models in parallel, and returning large model objects back, the shuffling of data can take a large chunk of that time. Rather than returning the entire model object, you might consider only returning a performance metric, or smaller specific pieces of that model that you are most interested in.
This performance drop can especially be prominent if using
future_pmap() to iterate over rows and return large objects at each iteration.