Benchmarking Single- and Multi-Core BLAS Implementations and GPUs for use with R

We provide timing results for common linear algebra subroutines across BLAS (Basic Linear Algebra Subprograms) and GPU (Graphics Processing Unit)-based implementations. Several BLAS implementations are compared. The first is the unoptimised reference BLAS which provides a baseline to measure against. Second is the Atlas tuned BLAS, configured for single-threaded mode. Third is the development version of Atlas, configured for multi-threaded mode. Fourth is the optimised and multi-threaded Goto BLAS. Fifth is the multi-threaded BLAS contained in the commercial Intel MKL package. We also measure the performance of a GPU-based implementation for R (R Development Core Team 2010a) provided by the package gputools (Buckner et al. 2010). Several frequently-used linear algebra computations are compared across BLAS (and LAPACK) implementations and via GPU computing: matrix multiplication as well as QR, SVD and LU decompositions. The tests are performed from an end-user perspective, and ‘net’ times (including all necessary data transfers) are compared. While results are by their very nature dependent on the hardware of the test platforms, a few general lessons can be drawn. Unsurprisingly, accelerated BLAS clearly outperform the reference implementation. Similarly, multi-threaded BLAS hold a clear advantage over single-threaded BLAS when used on modern multi-core machines. Between the multithreaded BLAS implementations, Goto is seen to have a slight advantage over MKL and Atlas. GPU computing is showing promise but requires relatively large matrices to outperform multi-threaded BLAS. We also provide the framework for computing these benchmark results as the corresponding R package gcbd. It enables other researchers to compute similar benchmark results which could be the basis for heuristics helping to select optimial computing strategies for a given platform, library, problem and size combination.