Algebra backends for R

Matrix operations in R are mostly executed by system-wide linear algebra library. Here I compare three different implementations: Blas, Atlas and OpenBlas. Benchmark is performed using microbenchmark 1.4 for R 3.4.4. All operations are carried out on matrix with 1000x1000 random entries which occupies approximetly 7.6Mb of RAM. Computations were performed on 8 core Intel's i7-6700HQ CPU @ 2.60GHz. OpenBlas and Atlas utilzied all 8 cores, whereas Blas run only on single core, as noted by htop.

Benchmark script accessing 4 common operations

library(microbenchmark)

x = matrix(runif(1000*1000),ncol=1000)
m = microbenchmark("inner" = { x%*%t(x) },
                   "crossprod" = {crossprod(x,x)},
                   "inverse" = {solve(x)},
                   "svd" = {svd(x)},
                   times = 100)

write.csv(m,'blas.txt')

I am using GNU/Linux Debian distribution, so backends were swapped by executing

update-alternatives --config libblas.so-x86_64-linux-gnu 
update-alternatives --config libblas.so.3-x86_64-linux-gnu
update-alternatives --config liblapack.so-x86_64-linux-gnu
update-alternatives --config liblapack.so.3-x86_64-linux-gnu

The following table shows 3 configurations tested

With sufficiently large matrix OpenBlas performs best in computing crossproduct, inner product, inverse and singular value decomposition (svd). Worst is Blas which is not parallelized. Inversion of the matrix with OpenBlas had some issues with high variablility and outliers, so I repeated run with OpenBlas two more times and outliers disapeared in third run.