I set out to find out whether I could speed up R by compiling it from source and:

• using Apple´s Accelerate Framework
• enabling OpenMP (which is disabled under OS X and Windows by default, but enabled under Linux)
• using Intel´s Intel´s Math Kernel Library

I also wanted to know how an implicit parallel library, like OpenMP, performs within explicit parallelism, e.g. calls from the parallel package.

So I compiled 6 different versions of R 3.01 on OS X and tested them for speed against the OS X .pkg from CRAN. I used gcc 4.8 (not the gcc 4.2 from Apple), which you can download from here. For MKL, I used both the GNU (gcc, gfortran) and the Intel compilers (icc, ifortran). The 6 versions and their configure settings are:

1. R 3.01 compiled with Apple´s Accelerate framework
2. R 3.01 compiled with OpenMP enabled
3. R 3.01 compiled with gcc and gfortran using Intel MKL (sequential)
4. R 3.01 compiled with icc and ifortran using Intel MKL (sequential)
5. R 3.01 compiled with gcc and gfortran using Intel MKL (threaded)
6. R 3.01 compiled with icc and ifortran using Intel MKL (threaded)

To measure the speed of the 7 Versions of R I now had (6 compiled, 1 .pkg from CRAN), I used Simon Urbanek´s R Benchmark 2.5. I let it execute serially for 8 runs, and then in parallel with 4 cores and 2 runs each (so also 8 in total). Additionally, I let each version of R carry out a large matrix multiplication (10000 rows and 5000 columns). Click here to view the benchmark script. I ran the tests on an Intel Core i5 clocked at 2.9 GHz with 16 GB RAM.

Elapsed time in seconds for various R builds and benchmarks under OS X:

So what do all the NAs mean? None of the R versions I compiled could execute mclapply() without crashing. If anyone knows how to fix this, please drop me a message. The benchmarks of the R versions that did run were much faster though. The Matrix multiplication was on average 2700% faster, and the more diversified R Benchmark was around 400% faster than stock R on OS X for the optimized libraries. Nevertheless, if I cannot fix the crashes that occur with the parallel library, I am going to stick with the stock R version from CRAN. Any suggestions to improve the compiler options or tests to add to the profiling script are very welcome.