multidplyr is a backend for dplyr that partitions a data frame across
multiple cores. You tell multidplyr how to split the data up with
partition() and then the data stays on each node until you explicitly
retrieve it with
collect(). This minimises the amount of time spent
moving data around, and maximises parallel performance. This idea is
inspired by partools by Norm
Matloff and distributedR by
the Vertica Analytics team.
Due to the overhead associated with communicating between the nodes, you won’t see much performance improvement with simple operations on less than ~10 million observations, and you may want to instead try dtplyr, which uses data.table. multidplyr’s strength is found parallelsing calls to slower and more complex functions.
(Note that unlike other packages in the tidyverse, multidplyr requires R 3.5 or greater. We hope to relax this requirement in the future.)
You can install the released version of multidplyr from CRAN with:
And the development version from GitHub with:
# install.packages("devtools") devtools::install_github("tidyverse/multidplyr")
To use multidplyr, you first create a cluster of the desired number of workers. Each one of these workers is a separate R process, and the operating system will spread their execution across multiple cores:
library(multidplyr) cluster <- new_cluster(4) cluster_library(cluster, "dplyr") #> #> Attaching package: 'dplyr' #> The following objects are masked from 'package:stats': #> #> filter, lag #> The following objects are masked from 'package:base': #> #> intersect, setdiff, setequal, union
There are two primary ways to use multidplyr. The first, and most efficient, way is to read different files on each worker:
# Create a filename vector containing different values on each worker cluster_assign_each(cluster, filename = c("a.csv", "b.csv", "c.csv", "d.csv")) # Use vroom to quickly load the csvs cluster_send(cluster, my_data <- vroom::vroom(filename)) # Create a party_df using the my_data variable on each worker my_data <- party_df(cluster, "my_data")
Alternatively, if you already have the data loaded in the main session,
you can use
partition() to automatically spread it across the workers.
partition(), it’s a good idea to call
ensure that all of the observations belonging to a group end up on the
library(nycflights13) flight_dest <- flights %>% group_by(dest) %>% partition(cluster) flight_dest #> Source: party_df [336,776 x 19] #> Groups: dest #> Shards: 4 [81,594--86,548 rows] #> #> year month day dep_time sched_dep_time dep_delay arr_time sched_arr_time #> <int> <int> <int> <int> <int> <dbl> <int> <int> #> 1 2013 1 1 544 545 -1 1004 1022 #> 2 2013 1 1 558 600 -2 923 937 #> 3 2013 1 1 559 600 -1 854 902 #> 4 2013 1 1 602 610 -8 812 820 #> 5 2013 1 1 602 605 -3 821 805 #> 6 2013 1 1 611 600 11 945 931 #> # … with 336,770 more rows, and 11 more variables: arr_delay <dbl>, #> # carrier <chr>, flight <int>, tailnum <chr>, origin <chr>, dest <chr>, #> # air_time <dbl>, distance <dbl>, hour <dbl>, minute <dbl>, time_hour <dttm>
Now you can work with it like a regular data frame, but the computations
will be spread across multiple cores. Once you’ve finished computation,
collect() to bring the data back to the host session:
flight_dest %>% summarise(delay = mean(dep_delay, na.rm = TRUE), n = n()) %>% collect() #> # A tibble: 105 x 3 #> dest delay n #> <chr> <dbl> <int> #> 1 ABQ 13.7 254 #> 2 AUS 13.0 2439 #> 3 BQN 12.4 896 #> 4 BTV 13.6 2589 #> 5 BUF 13.4 4681 #> 6 CLE 13.4 4573 #> 7 CMH 12.2 3524 #> 8 DEN 15.2 7266 #> 9 DSM 26.2 569 #> 10 DTW 11.8 9384 #> # … with 95 more rows
Note that there is some overhead associated with copying data from the
worker nodes back to the host node (and vice versa), so you’re best off
using multidplyr with more complex operations. See
vignette("multidplyr") for more details.