A few comments (I have done extensive comparisons in all cells of
S+/R x Linux/Windows).
1) There are lots of differences between Windows versions, and indeed
Linux versions (especially with the gcc version fiascos and the problems
with learning how to get optimal performance out of gcc 3.0.x).
2) There are lots of differences between S-PLUS versions, and indeed R
versions.
Given that, as R uses gcc on both Linux and Windows there is a consistent
picture: a given version of R is 5-20% slower on professional versions of
Windows (NT/2000/XP) than Linux on the same hardware. I find that S-PLUS
for Windows is usually closer to Linux, but as I have never had an
identical codebase on each it is hard to be precise: on some problems
S+6.0.3 (Windows) is a lot faster than 6.0.1 (linux).
My experience is that the name of the OS is one of the least significant
issues in determining performance. (The biggest one is probably the
the programmer: I have seen speedups from days to less than a
second, the first being a serious best attempt.)
BTW to S.D. Byers: awk, sed etc are POSIX not Unix tools, widely available
on Windows and other OSes, and perl has for ages (if not for ever) been
cross-platform.
On Fri, 22 Feb 2002, Paul, David A wrote:
> I have summarized the responses below. -david paul
>
>
> (1) Is Splus faster (on a PC) under Linux than under
> Windows? I am specifically thinking of intensive
> loops here, but would also solicit information
> regarding how Splus stores information within
> the Linux system vx. Windows.
>
> Response #1: Not significantly different, AFAIK. Linux
> has better process management than Windoze, so on Linux
> the Splus/R process need not hog all the resources
> in a single CPU system. -Andy Liaw
>
> Response #2: Everything is always faster on Linux,
> especially if you recompile the kernel for your specific
> machine. I think Linux ships compiled for a 386,
> but you can recompile for 486, Pentium class, etc.
> -Bruce McCullough
>
>
> (2) Can Splus under Linux take advantage of multiple
> processors? I know there was some discussion on
> the listserv earlier about multiple processors,
> but I don't recall if the Linux OS was mentioned.
>
> After being chastised for not looking in the archives, I
> realized that there WAS such a thing as archives and searched
> for answers to this question. It turns out that Splus cannot
> take advantage of multiple processors, but does receive some
> small benefit from running on a dual-processing machine since
> it effectively gets its own processor. The best bet is to
> increase the memory. However, my archive search yielded a
> result that the folks at Bell Labs were (in 1998) working on
> a multi-threaded version of Splus. There is a nice link to
> a dissertation that focuses on this specific issue. Don't
> print it when your co-workers are trying to print, though, as
> it is 170 pages long.
>
>
> (3) Is R capable of taking advantage of multiple processors
> under either the Windows or Linux OS?
>
> Response #1: Similar to Splus. Splus is linked with Intel
> Math Kernel Library for fast linear algebra operations.
> R can be compiled with ATLAS on both Linux and Windoze to achieve
> similar effect. I have compiled R with Intel MKL on Linux, and I
> do get both CPUs to work simultaneously, e.g., on matrix
> multiplication (but not much else). There's also a rpvm package
> for running R on clusters. -Andy Liaw
>
>
> (4) How does R compare to Splus with regards to (1) above?
>
> Response #1: R's for loop is *generally* faster and more memory
> efficient than Splus. I believe the gap has narrowed with Splus
> 6. It also depends a whole lot on what actually gets done inside
> the loop. R doesn't do *everything* faster than Splus. For some
> tasks R is actually quite a bit slower than Splus. -Andy Liaw
>
> Response #2: Personally I have started migrating to R under linux.
> It has come down to useability rather than speed, although speed
> is an issue. S+ under the linux is the complicated new stuff that
> does not gel with the simplicity I desire. I have avoided that
> since I bought it and tested it. I do not regard Splus or R as
> particularly usable under windows. Having other unix functionality
> sitting ready to go such as awk, perl, sed etc etc is the key for
> large powerful analyses. Doing some of the donkey work outside R/S+
> is useful in some cases and essential in others. That said my data
> are often rather large, like 250M lines etc etc. -S.D. Byers
>
> Response #3: I have done a number of my own comparisons of
> Windows vs Linux. My setup is that I have two harddrives, one for
> Linux and one for Windows. I have Linux and Windows versions of
> SPlus, Mathematica, RATS, TSP, and Tex/LaTeX. I did this experiment
> with Windows 95 and Red Hat 5.2 (so this was years ago): A couple
> programs I have run on TSP (Time Series Processor, a package for
> doing econometrics) with both Windows 2000 and Red Hat 7.1, and
> Linux is almost twice as fast. -Bruce McCullough
>
>
>
>
>
>
> --------------------------------------------------------------------
> This message was distributed by s-news@lists.biostat.wustl.edu. To
> unsubscribe send e-mail to s-news-request@lists.biostat.wustl.edu with
> the BODY of the message: unsubscribe s-news
>
--
Brian D. Ripley, ripley@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 272860 (secr)
Oxford OX1 3TG, UK Fax: +44 1865 272595
|