Beware Hidden Parallelisation!
Optimised Linear Algebra Libraries
BLAS (Basic Linear Algebra Subprograms) is a set of low-level routines for performing common linear algebra operations, such as vector-vector operations, matrix-vector operations, and matrix-matrix operations.
R uses BLAS (Basic Linear Algebra Subprograms) and LAPACK (Linear Algebra PACKage) libraries for linear algebra operations. Because the default libraries have not been fully optimized, you could see large increases in speed, for code that is dependent on linear algebra computations, by switching to a different BLAS library.
Optimised BLAS libraries
Optimized BLAS libraries are faster than unoptimized ones for several reasons, including:
C implementation: Optimized BLAS libraries are typically written in C. The use of C allows for efficient implementation of the basic linear algebra operations that are performed by BLAS libraries.
Vectorisation: Optimized BLAS libraries often use Vectorisation.
Parallel computation: Many optimized BLAS libraries are designed to take advantage of parallel computation. This allows for faster computation of the linear algebra operations, as the work is divided among multiple processors.
Advanced algorithms: Optimized BLAS libraries often implement advanced algorithms for linear algebra operations that are designed to reduce the number of operations required or to take advantage of specific hardware features, such as specialized vector processing units.
Examples of optimised BLAS libraries include OpenBLAS, Intel MKL (Math Kernel Library), and Atlas (Automatically Tuned Linear Algebra Software)
To use an optimized BLAS library in R, you typically need to install the library and configure R to use it. The exact steps for doing this will depend on the specific BLAS library you want to use and the operating system you are using.
Further Reading
OpenMP
OpenMP is a popular technology for adding parallelism to C, C++, and Fortran programs, and some BLAS libraries choose to implement parallelism using OpenMP to take advantage of multiple cores on a single machine.
Many R packages can take advantage of OpenMP to perform parallel computation under the hood, including data.table and fst which can run tasks like groub_by operations or reading/writing data in parallel.
It is important to note that the specific steps for installing an R package with support for OpenMP can vary depending on the package, the operating system, and the version of R, so it is always a good idea to consult the package documentation or ask the package maintainer for more detailed information.
If your’re trying to parallelise your R code while also using one of these optimised libraries which parallelise at the compiled language level, you may soon overwhelm any resources requested!
If you want to use parallel R code with one of these libraries/packages, there are environment variables you can use to switch off the parallelisation at the compiled library or openMP level.
If you want to use such libraries on HPC for fast execution on Linear algebra routines, don’t add additional parallelisation in R. Instead ask for a single task and multiple cores and let the optimised libraries take advantage of them.