On Thu, Jun 3, 2010 at 1:16 PM, Dirk Eddelbuettel <[email protected]> wrote: > >/. > > Besides Goto BLAS, you may also enjoy the (multithreaded-by-default) Intel > MKL Blas you can get via Ubuntu's 'revolution-mkl' package for amd64 and i386. > > Hth.
That does help! Thanks. I've compared it with GotoBLAS2, and they are both much better than the default. GotoBLAS2 is a bit faster, but I only ran a couple of test R programs to compare, and the difference is small. Yesterday, there was a post here about Atlas BLAS not using many cores. I have found that not all big programs in R cause multiple cores to be used, but here is an example that does. I've run this with the Intel MKL or GotoBLAS and I do see all cores light up. mm <- matrix(rnorm(10^7), ncol = 10^3) ### Following does not cause use of more cores system.time(crossprod(mm)) ### Following DOES cause use of more cores system.time(mmm <- mm %*% t(mm)) As usual, you get more questions from me in response. 1. Will an R call to BLAS automatically spread out across cores? 2. I'm reading a doc from Intel about MKL. It is quite understandable, even for cavemen like me :) http://www.ucl.ac.uk/research-computing/.../mkl_10.2_documentation.pdf It describes optimization with parallel BLACS and "scalapack". If I want to see that in action, though, I have to purchase the actual Intel product, right? With revolution-mkl, I get: /usr/lib/R/lib/libblas.so.3gf /usr/lib/R/lib/libguide.so /usr/lib/R/lib/liblapack.so.3gf Keep in mind I'm the guy who thinks LD_LIBRARY_PATH is "magic" when I ask this: To use Intel's scalapack, what else has to be purchased? Use of scalapack and Rscalapack pre-suppose a working MPI layer, such as "OpenMPI". And the Intel scalapack implementation is likely to require me to use their specific version of MPI. 3. I've downloaded the source packaging for that and studied the license. Do you think I'm reading this right? I am allowed to use those shared object library files on other Linux systems i administer? It says I can use those files in any way consistent with Ubuntu policy, as long as I don't try to decompile or reverse engineer them. Am I reading this too liberally? pj > > -- > Regards, Dirk > -- Paul E. Johnson Professor, Political Science 1541 Lilac Lane, Room 504 University of Kansas _______________________________________________ R-SIG-Debian mailing list [email protected] https://stat.ethz.ch/mailman/listinfo/r-sig-debian
