Quick and dirty branchmark

2 min read

benchmark naniar Performance

A few months ago, I played with Nicholas’s naniar 📦 with performance in mind.

In the mean time, I have been busy doing other things and Jim’s patch was merged instead. Fair enough, the patch is R-only and much simpler, mine was using parallel c++ with Rcpp and RcppParallel.

I spent some time revisiting the cpp-test branch this afternoon, well tbh most of the time was spent trying to please git rebase. rebase is amazing, but when there are lots of commits in each branch, it can take some time and a few git rebase --continue attempts. It’s worth learning about it though, it makes git history cleaner.

Once properly rebased, and with a few extra edits, I finally had a version in the cpp-test branch I wanted to compare with the current version (in the master branch). That’s the second time this week I’ve needed a tool to compare performance of two branches of some repo.

Thanks to the wonders of twitter, I know that I should take a look at the Rperform package. I have not taken that time yet, and used a pattern involving callr and withr for a quick and dirty branchmark (that’s a benchmark to test accross branches).


with_libpaths("timings/master", install_github( "njtierney/naniar", ref = "master" ) )
with_libpaths("timings/cpp-test", install_github( "njtierney/naniar", ref = "cpp-test" ) )

fun <- function(){
  d <- purrr::map_df(1:10000, ~airquality)
  print(system.time(res <- add_n_miss(d)))

r1 <- callr::r(fun, libpath = "timings/master"  , show = TRUE)
r2 <- callr::r(fun, libpath = "timings/cpp-test", show = TRUE)

So initially I install the two versions (master and cpp-test) of the naniar package in their own libraries, using with_libpaths, and then using callr::r to evaluate the code to benchmark against each of the versions.

That’s what happens on my 💻, a pretty decent macbook pro late 2017 equipped with an i7.

> r1 <- callr::r(fun, libpath = "timings/master"  , show = TRUE)
   user  system elapsed 
  0.183   0.036   0.219 
> r2 <- callr::r(fun, libpath = "timings/cpp-test", show = TRUE)
   user  system elapsed 
  0.013   0.010   0.004 
> identical(r1,r2)
[1] TRUE

Probably also worth noting that I have -O3 setup in my ~/.R/Makevars, so the C++ code is correctly optimised.

CXX11FLAGS = -Wno-unused-result -O3