By Avraham
With the recent releases of R 3.2.4 and OpenBLAS 2.17, I decided it was time to re-benchmark R speed. I’ve settled on a particular set of tests, based on my experience as well as some of Simon Urbanek’s work which I separated into two groups: those focusing on BLAS-heavy operations and those which do not. I’ve posted the code I use to its own page, but I’ll copy it below for convenience:
set.seed(4987) library(microbenchmark) library(Matrix) AMy machine is a i7-2600K overclocked to 4.65Ghz with 16GB RAM, running Win7 64. I've also used RStudio Version 0.99.893 for each of these tests.
I ran the tests on four versions of R:
- Plain vanilla R version 3.2.4 Revised (2016-03-16 r70336) installed from the binary on CRAN
- The same version of R as #1, but with Rblas.dll replaced by one based on OpenBLAS 2.17 compiled specifically for the SandyBridge CPU (as part of the #3)
- R compiled from source using the gcc 4.9.3 toolchain, passing
-march=nativeto EOPTS and linking to a multi-threaded (4 max) SandyBridge-specific OpenBLAS 2.17 - Same as #3, but activating link-time optimization for core R and the BLAS
This time I was a bit wiser, and saved the results to RDS objects so that I could combine them into one R session and compare them. I’ll post the raw output at the bottom of the page, but I think it’s more intuitive to look at graphical comparisons. The takeaway is if you can install a fast BLAS! The other optimization options had minor effects, sometimes better sometimes worse, but using a fast BLAS made a significant difference. Adding a Platform variable to the four sets of BLAS and NotBLAS outputs and then combining them in one session allowed for the generation of comparison plots. The call for these graphs, for those following along at home, is:
ggplot(BLAS) + geom_boxplot(aes(y = ...read moreSource:: r-bloggers.com