[R] My surprising experience in trying out REvolution's R

[David M Smith]

You can always turn the multi-threading off with:

setMKLthreads(1)

We tested your code with this setting, and it ran in exactly the same
time as CRAN R.

# David Smith

On Fri, Apr 24, 2009 at 7:34 AM, Jason Liao <JLIAO at hes.hmc.psu.edu> wrote:
> David, thanks for following this through.
>
> I do not know how big a matrix needs to be before the multi-core
> multi-threading will start save time. But it seems useful to build this
> protection in your distribution so that it will not do multi-core when
> multi-threading is more likely to do harm.
>
> Jason
> -----Original Message-----
> From: David M Smith [mailto:david at revolution-computing.com]
> Sent: Thursday, April 23, 2009 6:05 PM
> To: Jason Liao
> Cc: r-help at r-project.org
> Subject: Re: [R] My surprising experience in trying out REvolution's R
>
> We've taken a look at this in a bit more detail; it's a very
> interesting example.  Although the code uses several functions that
> exploit the parallel processing in REvolution R (notably %*% and
> chol), this was one of those situations where the overheads of
> threading pretty much balanced any performance gains: the individual
> matrices for the operations were too small.
>
> For some examples where the performance gains do show, see:
> http://www.revolution-computing.com/products/r-performance.php
>
> A more promising avenue for speeding up this code lies in
> parallelizing the outer for(i in 1:200) loop...
>
> # David Smith
>
> On Tue, Apr 21, 2009 at 9:26 AM, Jason Liao <JLIAO at hes.hmc.psu.edu>
> wrote:
>> I care a lot about R's speed. So I decided to give REvolution's R
>> (http://revolution-computing.com/) a try, which bills itself as an
>> optimized R. Note that I used the free version.
>>
>> My machine is a Intel core 2 duo under Windows XP professional. The
> code
>> I run is in the end of this post.
>>
>> First, the regular R 1.9. It takes 2 minutes and 6 seconds, CPU usage
>> 50%
>>
>> Next, REvolution's R. It takes 2 minutes and 10 seconds, CPU usage
> 100%.
>>
>>
>> In other words, REvolution's R consumes double the CPU with slightly
>> less speed.
>>
>> The above has been replicated a few times (as a statistician of
> course).
>>
>>
>> Anyone has any insight on this? Anyway, my high hope was dashed.
>
>
> --
> David M Smith <david at revolution-computing.com>
> Director of Community, REvolution Computing www.revolution-computing.com
> Tel: +1 (206) 577-4778 x3203 (San Francisco, USA)
>
> Check out our upcoming events schedule at
> www.revolution-computing.com/events
>



-- 
David M Smith <david at revolution-computing.com>
Director of Community, REvolution Computing www.revolution-computing.com
Tel: +1 (206) 577-4778 x3203 (San Francisco, USA)

Check out our upcoming events schedule at www.revolution-computing.com/events