[R-sig-hpc] Getting multithreaded R working (solution)
edd at debian.org
Fri Jul 30 00:16:25 CEST 2010
On 29 July 2010 at 23:46, Stefan Theussl wrote:
| > P.S. If anyone has *easy* instructions for getting atlas working as a
| > multithreaded library or how do drop in the Intel MKL (I couldn't
| > figure out which *.so would be the correct one) please let me know and
| > I'll post them as well.
| I did get MKL working with R via the following trick:
| 1) export
| 2) then configure with:
| "--with-blas=-L/opt/intel/intel-mkl-9.0.018/lib/em64t/ -lmkl -lguide
| Back then I used MKL version 9.0.018. I don't know if this still works
| with more recent MKL or R.
Step 2) should not be necessary. The BLAS is a binary interface; you can
"simply" drop the right ones in. By having ld.so "find" the MKL ones before
the system ones, we made the MKL be used in the Revolution R release with
Ubuntu 9.10. That still works for -- so no rebuilding of R needed here.
On 29 July 2010 at 14:56, Jonathan Greenberg wrote:
| Have you (or anyone) gotten Intel MKL working via the "plugin"
| approach that Dirk has setup for the debian packages, where you just
I have :)
I have a proof-of-concept shell script that takes advantage of the Debian /
Ubuntu package management system and does the following for benchmarks:
i) removes all accelerates blas and runs with 'refblas'
ii) installs Atlas and runs with 'atlas'
iii) removes Atlas, install MKL (via aforementioned Revo package) and
runs with 'mkl'
iv) remove MKL and install goto (using .deb package built with the neat
goto-helper package from the ISM in Japan) and run 'goto'
plus for kicks
v) install gputools and runs with 'gpu'
I showed first results in the 'Intro to HPC' tutorial at useR and want to do
some more work on this. In fact, the gputools authors and I have a mostly
empty project on r-forge for this but not had time to do much...
Dirk Eddelbuettel | edd at debian.org | http://dirk.eddelbuettel.com
More information about the R-sig-hpc