Improve parallel performance with batching in a static-dynamic branching pipeline - r

Question

BLUF: I am struggling to understand out how to use batching in the R targets package to improve performance in a static and dynamic branching pipeline processed in parallel using tar_make_future(). I presume that I need to batch within each dynamic branch but I am unsure how to go about doing that.
Here's a reprex that uses dynamic branching nested inside static branching, similar to what my actual pipeline is doing. It first branches statically for each value in all_types, and then dynamically branches within each category. This code produces 1,000 branches and 1,010 targets total. In the actual workflow I obviously don't use replicate, and the dynamic branches vary in number depending on the type value.
# _targets.r
library(targets)
library(tarchetypes)
library(future)
library(future.callr)
plan(callr)
all_types = data.frame(type = LETTERS[1:10])
tar_map(values = all_types, names = "type",
tar_target(
make_data,
replicate(100,
data.frame(x = seq(1000) + rnorm(1000, 0, 5),
y = seq(1000) + rnorm(1000, 20, 20)),
simplify = FALSE
),
iteration = "list"
),
tar_target(
fit_model,
lm(make_data),
pattern = map(make_data),
iteration = "list"
)
)
And here's a timing comparison of tar_make() vs tar_make_future() with eight workers:
# tar_destroy()
t1 <- system.time(tar_make())
# tar_destroy()
t2 <- system.time(tar_make_future(workers = 8))
rbind(serial = t1, parallel = t2)
## user.self sys.self elapsed user.child sys.child
## serial 2.12 0.11 25.59 NA NA
## parallel 2.07 0.24 184.68 NA NA
I don't think the user or system fields are useful here since the job gets dispatched to separate R processes, but the elapsed time for the parallel job takes about 7 times longer than the serial job.
I presume this slowdown is caused by the large number of targets. Will batching improve performance in this case, and if so how can I implement batching within the dynamic branch?

Answer 1

You are on the right track with batching. In your case, that is a matter of breaking up your list of 100 datasets into groups of, say, 10 or so. You could do this with a nested list of datasets, but that's a lot of work. Luckily, there is an easier way.
Your question is actually really well-timed. I just wrote some new target factories in tarchetypes that could help. To access them, you will need the development version of tarchetypes from GitHub:
remotes::install_github("ropensci/tarchetypes")
Then, with tar_map2_count(), it will be much easier to batch your list of 100 datasets for each scenario.
library(targets)
tar_script({
library(broom)
library(targets)
library(tarchetypes)
library(tibble)
make_data <- function(n) {
datasets_per_batch <- replicate(
100,
tibble(
x = seq(n) + rnorm(n, 0, 5),
y = seq(n) + rnorm(n, 20, 20)
),
simplify = FALSE
)
tibble(dataset = datasets_per_batch, rep = seq_along(datasets_per_batch))
}
tar_map2_count(
name = model,
command1 = make_data(n = rows),
command2 = tidy(lm(y ~ x, data = dataset)), # Need dataset[[1]] in tarchetypes 0.4.0
values = data_frame(
scenario = LETTERS[seq_len(10)],
rows = seq(10, 100, length.out = 10)
),
columns2 = NULL,
batches = 10
)
})
tar_make(reporter = "silent")
#> Warning message:
#> `data_frame()` was deprecated in tibble 1.1.0.
#> Please use `tibble()` instead.
#> This warning is displayed once every 8 hours.
#> Call `lifecycle::last_lifecycle_warnings()` to see where this warning was generated.
tar_read(model)
#> # A tibble: 2,000 × 8
#> term estimate std.error statistic p.value scenario rows tar_group
#> <chr> <dbl> <dbl> <dbl> <dbl> <chr> <dbl> <int>
#> 1 (Intercept) 17.1 12.8 1.34 0.218 A 10 10
#> 2 x 1.39 1.35 1.03 0.333 A 10 10
#> 3 (Intercept) 6.42 14.0 0.459 0.658 A 10 10
#> 4 x 1.75 1.28 1.37 0.209 A 10 10
#> 5 (Intercept) 32.8 7.14 4.60 0.00176 A 10 10
#> 6 x -0.300 1.14 -0.263 0.799 A 10 10
#> 7 (Intercept) 29.7 3.24 9.18 0.0000160 A 10 10
#> 8 x 0.314 0.414 0.758 0.470 A 10 10
#> 9 (Intercept) 20.0 13.6 1.47 0.179 A 10 10
#> 10 x 1.23 1.77 0.698 0.505 A 10 10
#> # … with 1,990 more rows
Created on 2021-12-10 by the reprex package (v2.0.1)
There is also tar_map_rep(), which may be easier if all your datasets are randomly generated, but I am not sure if I am overfitting your use case.
library(targets)
tar_script({
library(broom)
library(targets)
library(tarchetypes)
library(tibble)
make_one_dataset <- function(n) {
tibble(
x = seq(n) + rnorm(n, 0, 5),
y = seq(n) + rnorm(n, 20, 20)
)
}
tar_map_rep(
name = model,
command = tidy(lm(y ~ x, data = make_one_dataset(n = rows))),
values = data_frame(
scenario = LETTERS[seq_len(10)],
rows = seq(10, 100, length.out = 10)
),
batches = 10,
reps = 10
)
})
tar_make(reporter = "silent")
#> Warning message:
#> `data_frame()` was deprecated in tibble 1.1.0.
#> Please use `tibble()` instead.
#> This warning is displayed once every 8 hours.
#> Call `lifecycle::last_lifecycle_warnings()` to see where this warning was generated.
tar_read(model)
#> # A tibble: 2,000 × 10
#> term estimate std.error statistic p.value scenario rows tar_batch tar_rep
#> <chr> <dbl> <dbl> <dbl> <dbl> <chr> <dbl> <int> <int>
#> 1 (Inter… 37.5 7.50 5.00 0.00105 A 10 1 1
#> 2 x -0.701 1.17 -0.601 0.564 A 10 1 1
#> 3 (Inter… 21.5 9.64 2.23 0.0567 A 10 1 2
#> 4 x -0.213 1.55 -0.138 0.894 A 10 1 2
#> 5 (Inter… 20.6 9.51 2.17 0.0620 A 10 1 3
#> 6 x 1.40 1.79 0.783 0.456 A 10 1 3
#> 7 (Inter… 11.6 11.2 1.04 0.329 A 10 1 4
#> 8 x 2.34 1.39 1.68 0.131 A 10 1 4
#> 9 (Inter… 26.8 9.16 2.93 0.0191 A 10 1 5
#> 10 x 0.288 1.10 0.262 0.800 A 10 1 5
#> # … with 1,990 more rows, and 1 more variable: tar_group <int>
Created on 2021-12-10 by the reprex package (v2.0.1)
Unfortunately, futures do come with overhead. Maybe it will be faster in your case if you try tar_make_clustermq()?