R devel - Nov 1998 - [R] R functionality

> Date: Mon, 30 Nov 1998 12:07:10 -0500
> From: Paul Gilbert <pgilbert@bank-banque-canada.ca>
> To: Ed Kademan <kademan@phz.com>
> 
> >Is there anything R can do that Splus can't?
> 
> Problems which require lots of looping and cannot be re-coded as
vector/matrix
> operations are theoretically possible but practically infeasible in S,
unless
> they  are coded in C or Fortran and called from S. These problems seem to
run
> fine in R.
Some but not all. Last night I was trying to run my survival examples
in R on Windows, including a large loop or alternatively a vectorized
version using many 4000x3 matrices.  Both run happily in S-PLUS 4.5 in
about three minutes and less than 10Mb allocated. R needs a largish
heap (20Mb failed, 30Mb seemed enough) that got partially swapped out
(the machine has 32Mb RAM) and I gave up after an hour of thrashing on
the loop version, and the vectorized version seemed in far more
page-thrashing trouble. Eventually I found just the right heap size
(24Mb) to get the loop version to run in about ten minutes.  R also has
a large amount of code in memory by that stage in the computations (ca
300000 n cells used).

BTW there are enormous differences amongst S3, S-PLUS 3/4.x, S4 and
S-PLUS 5.x in this area.


From: Thomas Lumley <thomas@biostat.washington.edu>
> Some of the
> packages, because they are more recent than similar S(-PLUS) built-in
> features may be better designed (eg the tree packages from Terry Therneau
> and Brian Ripley,
To be fair, both of those went the other way, they were written as
S(-PLUS) addons and then ported (by me) to R.  The only example I am
aware of going from R to S is my (still unfinished) lqs package, which
started as a re-design of some ROBETH code and ended up much more
comprehensive than what was available for S-PLUS.
> As R is much faster on _some_ code than S-PLUS I have also done
> simulations in R that I could not do in S-PLUS. Whether this counts as an
> added feature is arguable, but it can certainly be useful.
Agreed, but as my anecdote shows, much slower on other code.  I have
done extensive timings on the same machine and I see only a very few
statistical jobs where the overall speed difference is more than two
either way (against S-PLUS 3.4/4.5: S-PLUS 5.0 is another ball game).

`Your mileage may differ.'


-- 
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

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

PG>Problems which require lots of looping and cannot be re-coded as
vector/matrix
PG> operations are theoretically possible but practically infeasible in S,
unless
PG> they  are coded in C or Fortran and called from S. These problems seem to
run
PG> fine in R.

BR>Some but not all...

TL> As R is much faster on _some_ code ...

BR>Agreed, but as my anecdote shows, much slower on other code.
BR>`Your mileage may differ.'

Well, for the record, my milage is that Splus 3.3 is generally about twice as
fast as R with Solaris 2.6 on a Sparc Ultra10 doing my standard test set of
problems. These tests call Fortran for the main loop of ARMA and state space
time series models and use Splus "For" loops (which are a real pain to
debug)
for monte-carlo simulations and "For" loops calling the
"For" loops (which are a
super pain to debug) for evaluating estimation techniques. In R, I use
"for"
loops.

However, I've been experimenting with some other time series models which I
have
not (and cannot easily) code in fortran. In R it is true that I have to tweak
the nsize and vsize. Too much will cause paging. But I can run the monte-carlo
simulations and also evaluate estimation techniques in a reasonable time. In
Splus it seems these would take years to run. Apart from setting the nsize and
vsize too large and causing paging I have not run into the situation that R is
orders of magnitude slower than Splus.

Paul Gilbert

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._