[R] R code for performance

[Peter Dalgaard]

Prof Brian Ripley <ripley at stats.ox.ac.uk> writes:

> On Mon, 6 Jun 2005 v.demartino2 at virgilio.it wrote:
> 
> > At office I'm cautiously introducing R to be used as the basic statistical
> > program, getting rid of licensed stuff or reducing the amount of it.
> > The aim of R would be to run generic statistical programs built & "consumed"
> > when needed and some static procedure dealing with time-series.
> > Now, we have substantially 3 OS platforms, win xp, linux and freebsd 5.4,
> > on similar PCs (pentium 4, 2-2.5 GHz). I have been asked by the boss to
> > test the "average" performance (in term of speed and memory use) of R on
> > each of this platform to stick with one of them on a couple of PCs.
> >
> > Could you please suggest an R source code (apart from the "static procedure"
> > I will obviously test) to be run on the three platforms to test performance?
> >
> > If there is nothing of the kind, any suggestion?
> 
> 'make check' runs a lot of R code and times it.  The tests for the
> stats package look most relevant to you.  Beware of simplistic
> 'benchmarks' that test code snippets not relevant to your usage (and
> that may apply to the R examples which tend to be small datasets).
> 
> We know Linux (non-R-shlib) outperforms Windows XP by ca 20%, and some
> comments I have seen here suggest it outperforms FreeBSD as well.  But
> are such differences enough to determine your choice?

The scripts from the MASS package can also be used as an informal
benchmark, perhaps a bit more of a realistic mix than the stats
package. (Or was there a reason that Brian didn't mention them?)

It might also be relevant to note that, at least for a while, there
isn't going to be a 64 bit Windows version (the compiler etc. tool
chain is missing) so if you have large memory requirements, Linux or
BSD is the way to go. They also tend to be much easier to get
configured for building your own packages or just for using C/Fortran
extensions. The flip side is of course the (perceived)
userfriendliness of Windows.

If you have hardcore linear algebra requirements (e.g. inversion of
large matrices), you need to look into builds linked against fast BLAS
code (Goto, ATLAS). Most of the standard builds do not use this, so
benchmarks will be quite misleading.

-- 
   O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)             FAX: (+45) 35327907