Skip to content

optimal hardware for computations in R?

5 messages · Amit Ghosh, Brian Ripley, Paul Y. Peng +1 more

#
Hi,

I am planning to buy a new PC for computing simulations in R under
Linux. I was searching the web/mailing list-archives for useful hints
about the "optimal" choice of hardware - surprisingly I found no recent
topics.

As far as I know, R doesn't use threads, so I think that there should be
no benefit in choosing a dual-processor machine.

So the remaining affordable choices seems to be Athlon XP, Pentium 4,
Xeon or Athlon64/Opteron. Are there any R-related benchmarks or should
one simply look about the "standard" benchmark-results (SPEC, etc.)? Any
hints or experiences would be appreciated!

Thanks in advance,

Amit
#
On Tue, 23 Mar 2004, Amit Ghosh wrote:

            
Most of seem to be buying dual Opterons, not least so we can potentially 
access more than 4Gb.
It certainly can use a threaded BLAS.  You can also do two simulation runs 
simultaneously (and surely you will be doing more than one run?).
It really does depend on what exactly your computations do.  There are R 
`benchmarks', but they are not typical tasks (for me, and probably for no 
one else).

I would buy either a dual Athlon MP or a dual Opteron, and not worry too
much about this -- anything you buy today will look slow next year, and
you are not likely to see differences as large as 2x on one processor.
#
I recently ordered a computer which is intended to run both WindowsXP
and Linux (of course both versions of R as well). Before placing the
order, I discussed it with our system managers. They highly recommanded
a system with one P4 CPU with Intel's so called "hyper-threading"
technology over a system with two CPU's, and they claimed that both OS's
can take benefits from the "hyper-threading" technology. I haven't got
the machine yet and don't know how fast it is. At least this is another
option available.

Paul.
Prof Brian Ripley wrote:
#
On Tue, 2004-03-23 at 08:08, Paul Y. Peng wrote:
<snip>

Paul,

Others may chime in here, but you should be aware that there are still
lingering problems with Linux and HT in uni-processor systems, at least
using stock 2.4 (and even 2.6 kernels).

There is a good article at 2cpu.com on HT and the 2.6 kernels here:

http://www.2cpu.com/articles/41_1.html

Under FC1, which I use, there are issues with the 2.4 SMP kernels with
HT enabled. I have a 3.2 Ghz P4 with HT in a Dell i5150 laptop. I had to
disable HT in BIOS and am running the UP kernel, due to a list of known
bugs in the FC1 2.4 kernel series, which include boot lockups, other
boot time errors and even things as subtle as keyboard related problems.

Also, according to Alan Cox at RH, there are still performance tuning
issues relative to process and thread scheduling for HT on the 2.6
kernels that are yet to be included (but will be).

As you will see from the above article, the gains to be had from HT are
likely to be situationally specific and not an "all or nothing" gain. 

HTH,

Marc Schwartz
#
On Tue, 23 Mar 2004, Paul Y. Peng wrote:

            
I do have such a machine as my home machine, so I already have experience
under Fedora Core 1 and Windows XP. The gain over a single processor is
small compared to a dual processor (at best 1.2x in my experience), and
many magazine tests have recommended turning hyperthreading off.  I know
that a dual Athlon 2600 (my office machine) compiles R about twice as fast
as a Pentium 2.6HT, for example.