[R] [OT] R on Atlas library

[Matthias Gondan]

Hi Allan,

Thank you for your response. Finally I succeeded. These were the 
commands to compile Atlas and R:

Atlas:

../configure -t 4 -Fa alg -fPIC
make
make install

(-fPIC is needed to wrap some of objects of lapack.a as into a shared lib)


R:
../configure --with-blas="/usr/local/lib/libptf77blas.a 
/usr/local/lib/libatlas.a -lpthread"
make
make install

Using your benchmark example, all CPUs seem to be used. It is now
working at such a high speed that I could not entirely be sure, "top"
refreshes only every 5 s.

Cheers

Matthias

Am 09.08.2010 11:51, schrieb Allan Engelhardt:
> I don't know about the specific function (a simple, stand-alone
> reproducible example is always helpful, see the posting guide for
> details), but ATLAS certainly uses all my cores on a test like this:
>
> s <- 7500 # Adjust for your hardware
> a <- matrix(rnorm(s*s), ncol = s, nrow = s)
> b <- crossprod(a) # Uses all cores here.
>
> My configuration of R with ATLAS on Linux (Fedora) is described at
> http://www.cybaea.net/Blogs/Data/Faster-R-through-better-BLAS.html
>
> Maybe your distribution has single-threaded ATLAS and you forgot to
> rebuild it with "enable_native_atlas 1" or the equivalent for your
> platform?
>
> Allan
>
>
> On 06/08/10 15:12, Matthias Gondan wrote:
>> Dear List,
>>
>> I am aware this is slightly off-topic, but I am sure there are people
>> who already had the problem and who perhaps solved it.
>>
>> I am running long-lasting model fits using constrOptim command. At work
>> there is a linux computer (Quad Core, debian) on which I already have
>> compiled R and Atlas, in the hope that things will go faster on that
>> machine.
>>
>> Atlas offers the possibility to be compiled for multiple cores, I used
>> that option, but without success. Performance meter ("top") for Linux
>> shows 25% CPU usage, and there is no loss of performance if I run
>> 4 instances of R doing heavy matrix multiplications. Performance drops
>> when a 5th instance of R is doing the same job, so I assume my attempt
>> was not successful.
>>
>> I am sure I did something wrong. Is there anybody who sucessfully
>> runs R with Atlas and all processors? A short description of the
>> necessary steps would be helpful.
>>
>> Searching around the internet was not very encourageing. Some people
>> wrote that it is not so simple to have Atlas fully exploit a multicore
>> computer.
>>
>> I hope this is not too unspecific.
>>
>> Best wishes,
>>
>> Matthias
>>
>

-- 
Dr. rer. nat. Matthias Gondan
Institut für Psychologie
Universität Regensburg
D-93050 Regensburg
Tel. 0941-943-3856
Fax 0941-943-3233
Email: matthias.gondan at psychologie.uni-regensburg.de
http://www.psychologie.uni-r.de/Greenlee/team/gondan/gondan.html