Improving OS X OpenMx Performance with Default libRblas
neale
Joined: 07/31/2009
We have found that the default blas (basic linear algebra system) libraries supplied with R for MAC OS X have poor performance. Specifically, the following three commands, executed as root in a Terminal.app window reduces OpenMx execution time by 75% in some cases (ymmv).
cd /Library/Frameworks/R.framework/Libraries
sudo rm libRblas.dylib
sudo ln -s libRblas.0.dylib libRblas.dylib
Note that these commands simply reconfigure the link, they don't actually delete any libraries. To *undo* these commands one could do the following:
cd /Library/Frameworks/R.framework/Libraries
sudo rm libRblas.dylib
sudo ln -s libRblas.vecLib.dylib libRblas.dylib
Exactly why the vecLib version of libRblas is so slow is not clear; this is a problem for the R or OS X developers to tackle. Whether similar issues exist on linux and windows systems, I don't know, but perhaps others could experiment and comment.
Umm, you don't need to do
Umm, you don't need to do this if you're running the openmx binary on OS X using R 2.10 or 2.11. I reconfigured the package to load the non veclib shared library. We were messing around with the shared libraries because of libgfortran anyway. So you only need to follow the instructions if you are building from source code or running R 2.9.
Log in or register to post comments
In reply to Umm, you don't need to do by mspiegel
Excellent! I might suggest
Excellent! I might suggest deleting the post, but perhaps some R guru will notice it and identify a better solution.
Log in or register to post comments
In reply to Umm, you don't need to do by mspiegel
if building from source, does
if building from source, does it need doing after each build?
Log in or register to post comments
In reply to if building from source, does by tbates
You should only have to run
You should only have to run these commands once. The routes will stay there as OpenMx updates.
You might have to redo it if you reinstall R or upgrade to a new version of R, though.
Log in or register to post comments