R Performance Optimization Exercises: 20 Practice Problems
Twenty practice problems on R performance: profiling, vectorization, data.table, byte-compile, Rcpp, memory. Hidden solutions.
RRun this once before any exercise
library(microbenchmark)
library(profvis)
library(bench)
library(data.table)
library(dplyr)
library(furrr)
library(future)
Exercise 1: Benchmark two approaches
Difficulty: Beginner.
Show solution
RInteractive R
microbenchmark::microbenchmark(
loop = { s <- 0; for (i in 1:1000) s <- s + i; s },
vec = sum(1:1000),
times = 50
)
Exercise 2: Vectorize a loop
Difficulty: Intermediate.
Show solution
RInteractive R
x <- 1:10000
# slow
slow <- function() { y <- numeric(length(x)); for (i in seq_along(x)) y[i] <- x[i]^2; y }
# fast
fast <- function() x^2
Exercise 3: Preallocate vs grow
Difficulty: Intermediate.
Show solution
RInteractive R
grow <- function(n) { x <- numeric(); for (i in 1:n) x <- c(x, i); x }
prealloc <- function(n) { x <- numeric(n); for (i in 1:n) x[i] <- i; x }
# prealloc is 100x+ faster for large n
Exercise 4: apply vs vectorized
Difficulty: Intermediate.
Show solution
RInteractive R
m <- matrix(rnorm(1000*100), nrow = 100)
microbenchmark::microbenchmark(
apply = apply(m, 1, sum),
vec = rowSums(m),
times = 50
)
Exercise 5: Profile a function
Difficulty: Advanced.
Show solution
RInteractive R
profvis::profvis({
x <- numeric()
for (i in 1:1000) x <- c(x, i^2)
})
Exercise 6: byte-compile
Difficulty: Advanced.
Show solution
RInteractive R
slow <- function(n) { s <- 0; for (i in 1:n) s <- s + i; s }
fast <- compiler::cmpfun(slow)
microbenchmark::microbenchmark(slow(1e4), fast(1e4), times = 50)
Exercise 7: data.table vs dplyr aggregation
Difficulty: Advanced.
Show solution
RInteractive R
dt <- data.table::as.data.table(mtcars)
microbenchmark::microbenchmark(
dplyr_way = dplyr::summarise(dplyr::group_by(mtcars, cyl), m = mean(mpg)),
dt_way = dt[, .(m = mean(mpg)), by = cyl],
times = 50
)
Exercise 8: Memory usage
Difficulty: Advanced.
Show solution
RInteractive R
m <- matrix(rnorm(1e6), 1000, 1000)
object.size(m) |> format(units = "MB")
Exercise 9: gc() to free memory
Difficulty: Beginner.
Show solution
RInteractive R
large <- numeric(1e7)
rm(large); gc()
Exercise 10: Avoid copy-on-modify with data.table
Difficulty: Advanced.
Show solution
RInteractive R
dt <- data.table::as.data.table(mtcars)
dt[, kpl := mpg * 0.425] # in-place, no copy
Exercise 11: Rcpp for hot loop
Difficulty: Advanced.
Show solution
RInteractive R
# Rcpp::cppFunction("double sum_cpp(NumericVector x){ double s=0; for(int i=0;i<x.size();i++) s+=x[i]; return s; }")
# sum_cpp(1:1000)
Exercise 12: tracemem demo
Difficulty: Advanced.
Show solution
RInteractive R
x <- 1:5
tracemem(x)
x[1] <- 99 # prints memory copy address
untracemem(x)
Exercise 13: bench package timing
Difficulty: Intermediate.
Show solution
RInteractive R
bench::mark(
v1 = sum(1:1e4),
v2 = { s <- 0; for (i in 1:1e4) s <- s + i; s },
check = TRUE
)
Exercise 14: Use which() instead of all-row filter
Difficulty: Intermediate.
Show solution
RInteractive R
v <- runif(1e6)
microbenchmark::microbenchmark(
v1 = v[v > 0.5],
v2 = v[which(v > 0.5)],
times = 20
)
Exercise 15: Avoid factors when not needed
Difficulty: Advanced.
Show solution
RInteractive R
# When reading strings that won't be used in models, stringsAsFactors = FALSE or use read_csv()
# In tibbles, strings stay as character by default. In data.frame pre-R 4.0, they became factors.
Exercise 16: rowsum() for fast group sums
Difficulty: Advanced.
Show solution
RInteractive R
rowsum(mtcars$mpg, mtcars$cyl)
Exercise 17: tabulate for fast counts
Difficulty: Advanced.
Show solution
RInteractive R
x <- sample(1:5, 1e5, replace = TRUE)
microbenchmark::microbenchmark(
table_fn = table(x),
tab = tabulate(x),
times = 20
)
Exercise 18: setDT in place
Difficulty: Intermediate.
Show solution
RInteractive R
df <- as.data.frame(mtcars)
data.table::setDT(df)
class(df)
Exercise 19: Faster reads with fread
Difficulty: Beginner.
Show solution
RInteractive R
data.table::fwrite(mtcars, "demo.csv")
data.table::fread("demo.csv")
Exercise 20: Parallel apply with future
Difficulty: Advanced.
Show solution
RInteractive R
future::plan(future::multisession, workers = 2)
furrr::future_map_dbl(1:5, ~ { Sys.sleep(0.1); .x^2 })
What to do next
- Parallel-Computing-Exercises (coming), parallel strategies.
- data.table-Exercises (shipped), performance via data.table.