[R-SIG-Mac] Broken BLAS

Jan de Leeuw deleeuw at stat.ucla.edu
Wed Sep 8 21:48:12 CEST 2010

This is more on post-multiplying a matrix by a diagonal matrix (in six different ways).

The machine-comparison results from this little exercise can be explained by three factors

-- CPU speed
-- number of threads
-- BLAS library

Of the six methods, the first two use BLAS, the last four do not. For the
last four, times are determined largely by CPU speed.

For the first two methods, there are two things going on. First, BLAS in the Windows
implementation of R is faster than the vecLib BLAS on OS X. Second, on the
Mac Pro with Nehalem, for reasons unknown (bug !!) the OS decides to use a non-parellel
single-threaded BLAS. Running the same code, with the same OS and the same R, on a MacBook
or an iMac or a MacAir, does use multiple threads.  This is all
under OS X 10.6.5, by the way, using R-develop from svn. Running the same code in Parallels
emulation, using Windows R on the MacPro, uses multiple threads as well.

I cannot possibly convey how annoying this is.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: matdiag.R
Type: application/octet-stream
Size: 1077 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-mac/attachments/20100908/a80c31d4/attachment.obj>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: imac.txt
URL: <https://stat.ethz.ch/pipermail/r-sig-mac/attachments/20100908/a80c31d4/attachment.txt>
-------------- next part --------------

     Jan de Leeuw, 11667 Steinhoff Rd, Frazier Park, CA 93225
     home 661-245-1725 mobile 661-231-5416 work 310-825-9550
     .mac: jdeleeuw +++  aim: deleeuwjan +++ skype: j_deleeuw
             I am I because my little dog knows me.
	 		                 Gertrude Stein

More information about the R-SIG-Mac mailing list