R help - Jan 2009 - R badly lags matlab on performance?

Here's a small R program:

---------------------------------------------------------------------------
a <- rep(1,10000000)

system.time(a <- a + 1)

system.time(for (i in 1:10000000) {a[i] <- a[i] + 1})
---------------------------------------------------------------------------

and here's its matlab version:

---------------------------------------------------------------------------
function forloop()

  a = ones(1e7,1);

  tic; a = a + 1; toc

  tic
  for i=1:1e7
    a(i) = a(i) + 1;
  end
  toc
---------------------------------------------------------------------------

The machine used for testing was an i386 core 2 duo machine at 2.2
GHz, running OS X `Tiger'. The R was 2.8.0 and the matlab was 2008a.

Here's the performance (time taken in seconds):

                        Matlab            R            Times
Vector version          0.0446            0.067        1.5x
For loop version        0.0992            42.209       425.5x

So the R is 1.5x costlier for the vector version and 425.5x costlier
with matlab.

I wonder what we're doing wrong!

-- 
Ajay Shah                                      http://www.mayin.org/ajayshah  
ajayshah at mayin.org                            
http://ajayshahblog.blogspot.com
<*(:-? - wizard who doesn't know the answer.

On Sat, 3 Jan 2009 22:25:38 +0530 Ajay Shah <ajayshah at mayin.org> wrote:

AS> system.time(for (i in 1:10000000) {a[i] <- a[i] + 1})

AS> I wonder what we're doing wrong!

it is no secret that R does badly with loops. Thats why it is
recommended to use vectorized operations.

Another approach is just in time compilation, which would speed up
simple loops. 

There is an interesting entry on R-wiki showing speeding up things:

http://wiki.r-project.org/rwiki/doku.php?id=tips:programming:code_optim2&s=just%20time

hth
Stefan

Ajay Shah <ajayshah <at> mayin.org> writes:
> 
> Here's a small R program:
> 
> ---------------------------------------------------------------------------
> a <- rep(1,10000000)
> 
> system.time(a <- a + 1)
> 
> system.time(for (i in 1:10000000) {a[i] <- a[i] + 1})
> ---------------------------------------------------------------------------
> 
> and here's its matlab version:
> 
> ---------------------------------------------------------------------------
> function forloop()
> 
>   a = ones(1e7,1);
> 
>   tic; a = a + 1; toc
> 
>   tic
>   for i=1:1e7
>     a(i) = a(i) + 1;
>   end
>   toc
> ---------------------------------------------------------------------------
> 
> The machine used for testing was an i386 core 2 duo machine at 2.2
> GHz, running OS X `Tiger'. The R was 2.8.0 and the matlab was 2008a.
> 
> Here's the performance (time taken in seconds):
> 
>                         Matlab            R            Times
> Vector version          0.0446            0.067        1.5x
> For loop version        0.0992            42.209       425.5x
> 
> So the R is 1.5x costlier for the vector version and 425.5x costlier
> with matlab.
> 
> I wonder what we're doing wrong!
  OK, I'll bite.

  1. are you sure that the difference in the vector version is
real?  On my computer the R system time for the
vector computation ranges from 0.052 to 0.1.
I would definitely check how much variation there is.

  2. Every language has different advantages and disadvantages.
I believe Matlab does a lot of byte-code compiling.  If you're
interested in this sort of thing, try Ra
<http://www.milbo.users.sonic.net/ra/index.html> , a version
of R that allows similar speed-ups.

  3. I appreciate that every language can be improved, but this
feels like something that is not "fixable" without changing
the language significantly.  You have lots of options if you need
to do these kinds of operations -- Matlab (if you can afford it),
Octave (does it run as quickly as Matlab?), programming in C
(see http://modelingwithdata.org/ ), coding the critical
non-vectorizable bits of your algorithm in C or FORTRAN, etc. etc.
etc..

  cheers
    Ben Bolker