Spedd: R vs S-plus

(Ted Harding) wrote:

Hi Folks,

Sorry to raise what has probably been discussed before,
but I an repeatedly struck by the comparative slowness
of S-plus for Windows compared with R for Linux when doing
much the same thing.

I don't have a direct comparison, because they're not
running on the same machine; but machine W has a faster
CPU and more RAM than machine L, yet S-plus on W seems
to take longer by quite a big factor (of the order of 5x)
than R on L.

My instincts say that "WIndows" is probably a significant
factor in the comnparison, but still ...

Prejudices?

Ideally, to compare R with S-plus, one should look at them
both on the same OS (Unix or Windows) on the same machine.

Can anyone give me clean comparative speeds?

It heavily depends on what you are doing, and on the versions of R and 
S-PLUS.

Yes, especially versions of S-PLUS. There are big differences between
recent versions of S-PLUS, and a comparison of 2000 vs 6.1 depends heavily
on the task.

It also depends on the version of Windows, and (especially in S-PLUS) the
file system type (NTFS/VFAT and even the versions of each) and if it is 
local or remotely mounted.

Given you do not link R against very specialized libraries such as ATLAS 
on one and not the other OS, I found no dramatic differences between R 
on Linux and Windows, but that might depend on the application as well.

I often compare on the same hardware, using Windows XP.  I'd say that on
average the Windows port is 10-20% slower (and we have some idea why), and
almost never 50% slower.

I don't think it is normal to see factors as large as 5 either way on real
tasks, provided there is a reasonable amount of RAM available.  (Both R
and S-PLUS under Windows run very slowly if there is a very small amount
of RAM.)  I used to keep extensive tables of the time taken for different
versions on the same hardware for all the MASS scripts, but these days
they run fast enough on all the systems I use.  Here's some numbers, RH8.0
on a dual Athlon 2600, R using ATLAS (single-processor)

	R 1.7.1	S+6.1
ch04	8.40	10.52
ch05	5.94	11.18
ch06	72.80	23.89
ch07	11.36	29.45
ch10	20.07	39.61
ch13	9.00	13.7

That's probably a fair comparison, as I have tried to make those tasks
work well on both systems *and* they are real tasks, not small artificial
`benchmarks'.

Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595