f3lix time spent with my computer

It is no news that R’s default BLAS is much slower that other available BLAS implementations. In A trick to speed up R matrix calculation/ Yu-Sung Su recommends using the ATLAS BLAS which is available on CRAN. When I learned about the possible speed-up  a while ago I tried several BLAS libraries and I found that GotoBLAS2 was giving me the best performance among the open-source BLAS implementations. Today I decided to check once again how much it makes sense to replace R’s default BLAS library.

Here are some results from my Intel i7-620M laptop running Windows 7:

I used rbenchmark for the measurements with the default of 100 repetitions and I saved and reused the same matrix across all runs:


A = matrix(rnorm(1000*1000),ncol=1000)
benchmark(A %*% A)
benchmark(svd(A))


A %*% A

	BLAS	seconds	speed-up
	MKL	31.65	7.16
	GotoBLAS2	42.30	5.36
	Default	226.53	1.00

svd(A)

	BLAS	seconds	speed-up
	MKL	317.61	12.52
	GotoBLAS2	389.13	10.22
	Default	3976.21	1.00

Conclusion

If you are using R operations that rely on BLAS than you should always use a faster BLAS implementation. If you are using Windows, you can use the Rblas.dll from here or you try the Revolution R Community edition which comes with a MKL BLAS.

Tags: BLAS

This entry was posted on Tuesday, October 30th, 2012 at 17:00 and is filed under R. You can follow any comments to this entry through the RSS 2.0 feed. You can leave a comment, or trackback from your own site.