Gardar Johannesson
2001-Dec-20 16:54 UTC
[R] optimizing R-1.4.0 build on Solaris; a show-and-tell storry
This is a little success story about the benefits of changing
the defaults in config.site when I was building R-1.4.0 for Solaris
(on a Sun Sparc that I'm currently using).
For previous versions of R, I had just used the default config.site and
not given it any thought. Since the Sun machine that I'm using
is not getting any faster, I decided I would give config.site a look
when building R-1.4.0.
By default, doing './configure' and then 'make' in building
R-1.4.0 from
source, results in the following (short summary list) of compiling flags:
BLAS = blas.o
BLAS_LIBS CC = cc
CFLAGS = -g
FC = f77
FFLAGS = -g
Following suggestions given in R-admin.html, I also build R-1.4.0
with:
BLAS BLAS_LIBS = -xlic_lib=sunperf -lsunmath
CC = cc -xarch=v9
CFLAGS = -xO5 -xlibmil -dalign
FC = f95 -xarch=v9
FFLAGS = -xO5 -xlibmil -dalign
I did few tests comparing the speed of these two builds. In short, I
saw about 65% speed improvement for general use, slightly more for
regression problems (2-3 times), and considerable more in matrix
multiplication (50 times).
Here are the tests.
1) Timing the tests/Examples/base-Ex.R script. I did the following for
the two builds:
time ./bin/R --vanilla < tests/Examples/base-Ex.R > tmp.out
resulting in the following times:
R-1.4.0-def: 227.70u 26.88s 4:20.34 97.7%
R-1.4.0-opt: 138.75u 30.90s 2:57.62 95.5%
for the default and optimized version, where 227.70u and 138.75u are
the users CPU time. That is, the default is about 65% slower.
2) A little MCMC example that I have using a for-loop to generate 10,000
samples from the posterior:
R-1.4.0-def: 14.45 sec user CPU
R-1.4.0-opt: 8.96 sec user CPU
S-6.0 : 34.19 sec user CPU
where the last line is from S-plus 6.0 on the same machine.
3) A regression,
lm(ozone ~ ns(lat.band,df=15) +
ns(lat.band,df=10):ns(lon.band,df=15),
weights=1/var, data=data, na.action=na.omit))
where data has in one case 3240 rows and in a other case 12960 rows.
The number of estimated parameters is 166 in both cases.
For data with 3240 rows:
R-1.4.0-def: 7.12 sec user CPU time
R-1.4.0-opt: 2.90 sec user CPU time
S-6.0 : 3.78 sec user CPU time
For data with 12960 rows:
R-1.4.0-def: 28.34 sec user CPU time
R-1.4.0-opt: 14.97 sec user CPU time
S-6.0 : 13.70 sec user CPU time
4) The result of system.time(B <- A %*% A) where A is 500x500 matrix.
R-1.4.0-def: 18.83 sec user CPU time
R-1.4.0-opt: 0.37 sec user CPU time
I hope this will be of use to somebody... cheers, Gardar
_________________________________________________________
Gardar Johannesson
Department of Statistics
Ohio State University
304E Cockins Hall, 1958 Neil Av.
Columbus, OH 43210
Tel: 614-292-1567
Fax: 614-292-2096
e-mail: gardar at stat.ohio-state.edu
WWW: www.stat.ohio-state.edu
_________________________________________________________
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help 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-help-request at
stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
Prof Brian Ripley
2001-Dec-20 18:17 UTC
[R] optimizing R-1.4.0 build on Solaris; a show-and-tell storry
That's a 64-bit build. Unless you really need it, -xarch=v8 (or omit it altogether) will be faster (about 10% on my tests). Most of the difference is turning optimization on (and debugging off) On Thu, 20 Dec 2001, Gardar Johannesson wrote:> This is a little success story about the benefits of changing > the defaults in config.site when I was building R-1.4.0 for Solaris > (on a Sun Sparc that I'm currently using). > > For previous versions of R, I had just used the default config.site and > not given it any thought. Since the Sun machine that I'm using > is not getting any faster, I decided I would give config.site a look > when building R-1.4.0. > > By default, doing './configure' and then 'make' in building R-1.4.0 from > source, results in the following (short summary list) of compiling flags: > > BLAS = blas.o > BLAS_LIBS > CC = cc > CFLAGS = -g > FC = f77 > FFLAGS = -g > > Following suggestions given in R-admin.html, I also build R-1.4.0 > with: > > BLAS > BLAS_LIBS = -xlic_lib=sunperf -lsunmath > CC = cc -xarch=v9 > CFLAGS = -xO5 -xlibmil -dalign > FC = f95 -xarch=v9 > FFLAGS = -xO5 -xlibmil -dalign > > I did few tests comparing the speed of these two builds. In short, I > saw about 65% speed improvement for general use, slightly more for > regression problems (2-3 times), and considerable more in matrix > multiplication (50 times). > > Here are the tests. > > 1) Timing the tests/Examples/base-Ex.R script. I did the following for > the two builds: > time ./bin/R --vanilla < tests/Examples/base-Ex.R > tmp.out > resulting in the following times: > R-1.4.0-def: 227.70u 26.88s 4:20.34 97.7% > R-1.4.0-opt: 138.75u 30.90s 2:57.62 95.5% > for the default and optimized version, where 227.70u and 138.75u are > the users CPU time. That is, the default is about 65% slower. > > 2) A little MCMC example that I have using a for-loop to generate 10,000 > samples from the posterior: > R-1.4.0-def: 14.45 sec user CPU > R-1.4.0-opt: 8.96 sec user CPU > S-6.0 : 34.19 sec user CPU > where the last line is from S-plus 6.0 on the same machine. > > 3) A regression, > lm(ozone ~ ns(lat.band,df=15) + > ns(lat.band,df=10):ns(lon.band,df=15), > weights=1/var, data=data, na.action=na.omit)) > where data has in one case 3240 rows and in a other case 12960 rows. > The number of estimated parameters is 166 in both cases. > For data with 3240 rows: > R-1.4.0-def: 7.12 sec user CPU time > R-1.4.0-opt: 2.90 sec user CPU time > S-6.0 : 3.78 sec user CPU time > For data with 12960 rows: > R-1.4.0-def: 28.34 sec user CPU time > R-1.4.0-opt: 14.97 sec user CPU time > S-6.0 : 13.70 sec user CPU time > > 4) The result of system.time(B <- A %*% A) where A is 500x500 matrix. > R-1.4.0-def: 18.83 sec user CPU time > R-1.4.0-opt: 0.37 sec user CPU time > > > I hope this will be of use to somebody... cheers, Gardar > > > _________________________________________________________ > Gardar Johannesson > Department of Statistics > Ohio State University > 304E Cockins Hall, 1958 Neil Av. > Columbus, OH 43210 > > Tel: 614-292-1567 > Fax: 614-292-2096 > e-mail: gardar at stat.ohio-state.edu > WWW: www.stat.ohio-state.edu > _________________________________________________________ > > > -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- > r-help 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-help-request at stat.math.ethz.ch > _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._ >-- 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 272860 (secr) Oxford OX1 3TG, UK Fax: +44 1865 272595 -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- r-help 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-help-request at stat.math.ethz.ch _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._