Using the Rcpp Timer

Dirk Eddelbuettel — written Jan 6, 2013 — updated Dec 30, 2016 — source

Sine the 0.10.2 release, Rcpp contains an internal class Timer which can be used for fine-grained benchmarking. Romain motivated Timer in a post to the mailing list where Timer is used to measure the different components of the costs of random number generation.

A slightly modified version of that example follows below.

#include <Rcpp.h>
#include <Rcpp/Benchmark/Timer.h>

using namespace Rcpp;

// [[Rcpp::export]]
NumericVector useTimer() {
    int n = 1000000;

    // start the timer
    Timer timer;
    timer.step("start");        // record the starting point

    for (int i=0; i<n; i++) {
    timer.step("get/put");      // record the first step

    for (int i=0; i<n; i++) {
        rnorm(10, 0.0, 1.0);
    timer.step("g/p+rnorm()");  // record the second step

    for (int i=0; i<n; i++) {
        // empty loop
    timer.step("empty loop");   // record the final step

    NumericVector res(timer);   // 
    for (int i=0; i<res.size(); i++) {
        res[i] = res[i] / n;
    return res;

We get the following result, each expressing the cost per iteration in nanoseconds, both cumulative (default) and incrementally (by taking differences).

res <- useTimer()
res          # normal results: cumulative
      start     get/put g/p+rnorm()  empty loop 
   0.000156 1571.057133 3920.082384 3920.085223 
diff(res)    # simple difference
    get/put g/p+rnorm()  empty loop 
1571.056977 2349.025251    0.002839 

The interesting revelation is that repeatedly calling GetRNGstate() and PutRNGstate() can amount to about 60% of the cost of RNG draws. Luckily, we usually only have to call these helper functions once per subroutine called from R (rather than repeatedly as shown here) so this is not really a permanent cost to bear when running simulations with R.

It also show the usefulness of a fine-grained timer at the code level.

tags: benchmark  rng 

Related Articles