[R] Which Linux OS on Athlon amd64, to comfortably run R?

Prof Brian Ripley ripley at stats.ox.ac.uk
Thu Dec 6 08:07:09 CET 2007


On Thu, 6 Dec 2007, Emmanuel Charpentier wrote:

> Prof Brian Ripley a écrit :
>> Note that Ottorino has only 1GB of RAM installed, which makes a 64-bit
>> version of R somewhat moot.  See chapter 8 of
>>
>> http://cran.r-project.org/doc/manuals/R-admin.html
>
> Thank you for this reminder|tip ! I didn't re-read this document since
> ... oh, my  ... very early (1.x ?) versions. At which time my favorite
> peeve against R was the fixed memory allocation scheme.
>
> I would have thought that 64 bits machines could take advantage of a
> wider bus and (marginally ?) faster instructions to balance larger
> pointers. Am I mistaken ?

Yes, it is more complex than that.  If you run 32-bit instructions on a 
x86_64, the physical bus is the same as when you run 64-bit instructions. 
The larger code usually means the CPU caches spill more often, and some 
64-bit chips have more 32-bit than 64-bit registers which allows better 
scheduling.

The R-admin manual reports on some empirical testing.  But when you have 
limited RAM the larger code and data for a 64-bit build will cause more 
swapping and that is likely to dominate performance issues on large 
problems.

Note that the comparisons depend on both the chip and the OS: it seems 
that on Mac OS 10.5 on a Core 2 Duo the 64-bit version is faster (on small 
examples).  The original enquiry was about 'amd64 linux', but I've checked 
Intel Core 2 Duo as well: on my box 64-bit builds are faster than 32-bit 
ones, whereas the reverse is true for Opterons.  So it seems that the 
architectural differences of Core 2 Duo vs AMD64 do affect the issue.

-- 
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


More information about the R-help mailing list