I was finally able to install Microsoft R Open on a machine that features Ubuntu Server 16.04 so now let's test!

Microsoft© website states that *"From its inception, R was designed to use only a single thread (processor) at a time. Even today, R works that way unless linked with multi-threaded BLAS/LAPACK libraries."* That's true.

But also they state that *"Compared to open source R, the MKL offers significant performance gains, particularly on Windows."*. Microsoft R open is, to the date, the best option to install R on Windows in a painless way. I'm not so sure about the significant performance gains, specially because their website does not show different results on different hardware under a controlled experiment. So, if you are going to talk about significant differences please respect statistics.

This post complements an older post. Like my referred post, this article does not present evidence to close the long discussion on benchmarks nor provides a careful test on different machines.

The benchmarks presented here are an adaptation of Simon Urbanek's benchmarks, and I'll be using the `version.compare`

package that is used for the results exposed in Microsoft's website.

It is important to notice that I did repeat each test 100 times and the results presented here correspond to the median of the repetitions of each test. This was made after a suggestion from George Vega Yon.

This is how the test was made:

- System: Ubuntu Server 16.04
- Processor: Intel© Xeon 2.27GHz (using eight cores)
- Memory: 32 GB (four DDR3 cards of eight gigabytes each)
- R version: 3.4.2 different compilations

Operations Median Time | R built with Intel MKL | Microsoft R Open |
---|---|---|

Cholesky factorization | 0.8190 | 0.8270 |

Linear discriminant analysis | 2.9425 | 3.1610 |

Matrix multiplication | 4.2075 | 4.2745 |

Principal component analysis | 4.6750 | 4.7980 |

QR decomposition | 6.1480 | 5.9570 |

Singular value decomposition | 4.1445 | 4.3415 |

Operations Median Time | R built with OpenBLAS | Microsoft R Open |
---|---|---|

Cholesky factorization | 0.8960 | 0.8260 |

Linear discriminant analysis | 3.0875 | 3.1625 |

Matrix multiplication | 4.0005 | 4.2675 |

Principal component analysis | 6.0795 | 4.7720 |

QR decomposition | 5.7610 | 5.9560 |

Singular value decomposition | 6.4725 | 4.3335 |

These partial results show that CRAN R (aka R Open Source) built with Intel MKL numerical libraries performs better that Microsoft R Open. But, how better? This can't be extrapolated to other systems but Ubuntu.

For some operations R augmented with alternative math kernel (MKL in this case) performs better than Microsoft R Open.

I used this script to obtain the results: