Strategies to Speedup R Code

By Selva Prabhakaran


(This article was first published on DataScience+, and kindly contributed to R-bloggers)

The for-loop in R, can be very slow in its raw un-optimised form, especially when dealing with larger data sets. There are a number of ways you can make your logics run fast, but you will be really surprised how fast you can actually go.
This posts shows a number of approaches including simple tweaks to logic design, parallel processing and Rcpp, increasing the speed by orders of several magnitudes, so you can comfortably process data as large as 100 Million rows and more.

Lets try to improve the speed of a logic that involves a for-loop and a condition checking statement (if-else) to create a column that gets appended to the input data frame (df). The code below creates that initial input data frame.

# Create the data frame

The logic we are about to optimise:
For every row on this data frame (df), check if the sum of all values is greater than 4. If it is, a new 5th variable gets the value “greater_than_4”, else, it gets “lesser_than_4”.

# Original R code: Before vectorization and pre-allocation
  for (i in 1:nrow(df)) { # for every row
    if ((df[i, 'col1'] + df[i, 'col2'] + df[i, 'col3'] + df[i, 'col4']) > 4) { # check if > 4
      df[i, 5] 

All the computations below, for processing times, were done on a MAC OS X with 2.6 Ghz processor and 8GB RAM.

Vectorise and pre-allocate data structures

Always initialize your data structures and output variable to required length and data type before taking it to loop for computations. Try not to incrementally increase the size of your data inside the loop. Lets compare how vectorisation improves speed on a range of data sizes from 1000 to 100,000 rows.

# after vectorization and pre-allocation
output  4) {

Raw Code Vs With vectorisation:

Take statements that check for conditions (if statements) outside the loop

Taking the condition checking outside the loop the speed is compared against the previous version that had vectorisation alone. The tests were done on dataset size range from 100,000 to 1,000,000 rows. The gain in speed is again dramatic.

# after vectorization and pre-allocation, taking the condition checking outside the loop.
output  4  # condition check outside the loop
  for (i in 1:nrow(df)) {
    if (condition[i]) {

Condition Checking outside loops:

Run the loop only for True conditions

Another optimisation we can do here is to run the loop only for condition cases that are ‘True', by initialising (pre-allocating) the default value of output vector to that of ‘False' state. The speed improvement here largely depends on the proportion of ‘True' cases in your data.

The tests compared the performance of this against the previous case (2) on data size ranging from 1,000,000 to 10,000,000 rows. Note that we have increase a ‘0' here. As expected there is a consistent and considerable improvement.

output  4
  for (i in (1:nrow(df))[condition]) {  # run loop only for true conditions
    if (condition[i]) {

Running Loop Only On True Conditions:

Use ifelse() whenever possible

You can make this logic much simpler and faster by using the ifelse() statement. The syntax is similar to the if function in MS Excel, but the speed increase is phenomenal, especially considering that there is no vector pre-allocation here and the condition is checked in every case. Looks like this is going to be a highly preferred option to speed up simple loops.

  output  4, "greater_than_4", "lesser_than_4")

True conditions only vs ifelse:

Using which()

By using which() command to select the rows, we are able to achieve one-third the speed of Rcpp.

# Thanks to Gabe Becker
  want = which(rowSums(df) > 4)
  output = rep("less than 4", times = nrow(df))
  output[want] = "greater than 4"
# nrow = 3 Million rows (approx)
   user  system elapsed 
  0.396   0.074   0.481 

Use apply family of functions instead of for-loops

Using apply() function to compute the same logic and comparing it against the vectorised for-loop. The results again is faster in order of magnitudes but slower than ifelse() and the version where condition checking was done outside the loop. This can be very useful, but you will need to be a bit crafty when handling complex logic.

# apply family
  myfunc  4) {
    } else {

apply function Vs For loop in R:

Use byte code compilation for functions cmpfun() from compiler package, rather than the actual function itself

This may not be the best example to illustrate the effectiveness of byte code compilation, as the time taken is marginally higher than the regular form. However, for more complex functions, byte-code compilation is known to perform faster. So you should definitely give it a shot.

# byte code compilation

apply vs for-loop vs byte code compiled functions:

Use Rcpp

Lets turn this up a notch. So far we have gained speed and capacity by various strategies and found the most optimal one using the ifelse() statement. What if we add one more zero? Below we execute the same logic but with Rcpp, and with a data size is increased to 100 Million rows. We will compare the speed of Rcpp to the ifelse() method.

system.time (output 

Below is the same logic executed in C++ code using Rcpp package. Save the code below as “MyFunc.cpp” in your R session's working directory (else you just have to sourceCpp from the full filepath). Note: the // [[Rcpp::export]] comment is mandatory and has to be placed just before the function that you want to execute from R.

// Source for MyFunc.cpp
using namespace Rcpp;
// [[Rcpp::export]]
CharacterVector myFunc(DataFrame x) {
  NumericVector col1 = as(x["col1"]);
  NumericVector col2 = as(x["col2"]);
  NumericVector col3 = as(x["col3"]);
  NumericVector col4 = as(x["col4"]);
  int n = col1.size();
  CharacterVector out(n);
  for (int i=0; i 4){
      out[i] = "greater_than_4";
    } else {
      out[i] = "lesser_than_4";
  return out;

Rcpp speed performance against ifelse:

Use parallel processing if you have a multicore machine

Parallel processing:

# parallel processing
cl  4
# parallelization with vectorization

Remove variables and flush memory as early as possible

Remove objects rm() that are no longer needed, as early as possible in code, especially before going in to lengthy loop operations. Sometimes, flushing gc() at the end of each iteration with in the loops can help.

Use data structures that consume lesser memory

Data.table() is an excellent example, as it reduces the memory overload which helps to speed up operations like merging data.

dt  4) {
      dt[i, col5:="greater_than_4"]  # assign the output as 5th column
    } else {
      dt[i, col5:="lesser_than_4"]  # assign the output as 5th column

Dataframe Vs Data.Table:

Speed Summary

Method: Speed, nrow(df)/time_taken = n rows per second
Raw: 1X, 120000/140.15 = 856.2255 rows per second (normalised to 1)
Vectorised: 738X, 120000/0.19 = 631578.9 rows per second
True Conditions only: 1002X, 120000/0.14 = 857142.9 rows per second
ifelse: 1752X, 1200000/0.78 = 1500000 rows per second
which: 8806X, 2985984/0.396 = 7540364 rows per second
Rcpp: 13476X, 1200000/0.09 = 11538462 rows per second

The numbers above are approximate and are based in arbitrary runs. The results are not calculated for data.table(), byte code compilation and parallelisation methods as they will vary on a case to case basis, depending upon how you apply it.

To leave a comment for the author, please follow the link and comment on their blog: DataScience+. offers daily e-mail updates about R news and tutorials on topics such as: Data science, Big Data, R jobs, visualization (ggplot2, Boxplots, maps, animation), programming (RStudio, Sweave, LaTeX, SQL, Eclipse, git, hadoop, Web Scraping) statistics (regression, PCA, time series, trading) and more…

Source:: R News

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.