[R-sig-hpc] Recommended Rig for $/MFlops, Linear Algebra
ivowel at gmail.com
Mon Mar 30 14:32:06 CEST 2009
I am still trying to find out how AMD and Intel processors compare on
linear algebra in R. R's work is such task-specific (ok, say 2 or 3
tasks) that this should be easily quantifiable. I would in particular
like to know whether the reasonably standard versions (OSX
precompiled; linux with Atlas) are bound by the speed of the SSE*
instructions, or by the speed of the floating point instructions at
this point. finally, I think it would be useful to learn which
standard benchmark in the literature (specfloat or scimark or ...)
would seem to be a good proxy for relative speed of processors.
if you know of such a piece, please point me to it. if no one else
has written such a piece, I would be happy to donate remote access to
an OSX machine, an Intel 2 core duo, and a phenom II to do some of it.
I do not know enough about the internals of R well enough so as to
write on the subject.
On Mon, Mar 30, 2009 at 3:41 AM, Markus Schmidberger
<schmidb at ibe.med.uni-muenchen.de> wrote:
> I can not give any information about the best architecture for R. But there
> is some great work going on for R and multiprocessors:
> Attached some first results from the "National Supercomputer HLRB II - SGI
> Altix 4700" in Munich. It is the computation time for a simple (but enough
> computation time) bootstrap example. You can see a nearly linear speed-up up
> to 64 processors!
> I hope to present more results (on more processors) at the UseR conference.
> ivowel at gmail.com schrieb:
>> dear readers---following dirk's advice on r-help, let me ask if someone
>> else has recently done research on what the top performing R platform for
>> the money is. My needs are for monte-carlo simulations (which therefore are
>> easily parallelizable to multiple cores; I don't think R can use "core i7"
>> threads), mostly linear algebra. I thought I would share what little
>> knowledge I gathered today (and collect any ideas/suggestions that I may
>> have overlooked).
>> the prime candidates would seem to be a dual Shanghai 2.3GHz system
>> ($175/processor) + $100/motherboard + DDR2 Reg RAM; or a single core i7 920
>> 2.66GHz ($280/processor) + $300/motherboard + DDR3 RAM. These two seem to
>> be heads and shoulders above their siblings in $/performance. The Intel rig
>> is about 20% more expensive...but being faster in linear non-parallel tasks
>> is nice, too.
>> In some benchmarks, Whetstone (on Tom's hardware) in particular, the core
>> i7 is almost twice as fast. More importantly, I found web benchmarks of 160
>> specfloat for a pair of 2.8GHz Core i7 Xeons vs. 105 for a dual Shanghai
>> 2.7. Extrapolating, this would mean that the dual Shanghai 2.3GHz system
>> does around 90, while the single core i7 920 does about 80 (including a
>> slight bonus for single-processor use). It also presumes, as I believe,
>> that specfloat is a good indication of how fast linear algebra will perform
>> on R (and Stata). my big unknown here is whether specfloat can use
>> multi-threading while R and Stata cannot or vice-versa. Does anybody know?
>> So, AMD= $5/Specfloat. Intel= $6.75/Specfloat. This is only
>> CPU/motherboard. Other components will cost similarly, and thus bring the
>> relative ratio of the two closer to one another.
>> I hope this helps. of course, in about a month, it will be different
>> [[alternative HTML version deleted]]
>> R-sig-hpc mailing list
>> R-sig-hpc at r-project.org
> Dipl.-Tech. Math. Markus Schmidberger
> Ludwig-Maximilians-Universität München
> IBE - Institut für medizinische Informationsverarbeitung,
> Biometrie und Epidemiologie
> Marchioninistr. 15, D-81377 Muenchen
> URL: http://www.ibe.med.uni-muenchen.de Mail: Markus.Schmidberger [at]
> Tel: +49 (089) 7095 - 4497
More information about the R-sig-hpc