[Rd] FFT

[Prof Brian Ripley]

On Fri, 7 Oct 2005, Jennifer Lai wrote:

>    Is FFT implemented in R takes advantage of multi-processors?
> I ran this benchmark from from http://www.sciviews.org, and AMD Opteron
> 2.2 GHz performs better than AMD Opteron 1.8 GHz on all test cases,
> except FFT operation.
> Both machines run same OSs (RedHat WS 3) and 2.2 GHz has more memory (2
> GB RAM) than 1.8 GHz (1 GB RAM). The only difference is that 1.8 GHz is
> a dual-processor machine, and
> 2.2 GHz is a single processor machine. Could this be the reason?
> Does anyone has insights on this?

I am pretty confident that nothing in Unix-alike R is multi-threaded. 
(The Windows port is.) However, you can use some multi-threaded addons, 
e.g. a BLAS library.

AFAIK those chips differ in more than just speed, including exact cache 
details.  Other details of the CPU architecture are often more important 
than small differences in clock speed.  Even small compiler differences 
can affect performance a lot.

I would also caution aginst drawing any conclusions at all from this sort 
of `benchmark' (indeed, would say it was a misuse of the term). 
Fortunately the computer world has moved on to using realistic tasks as 
comparators.  When we last had a large procurement (a many-processor 
system fufilled by Opterons) we used real data-analysis programs written 
in C++ and Java.  The tenders managed a very surprising difference in 
throughput on apparently identical CPUs, and results even from the same 
tender on different models of Opteron did not scale even approximately 
with clock speed.  In the end we learnt more about the support skills of 
the potential suppliers than about the optimized performance of their 
systems.

-- 
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