[R-sig-ME] How can I optimize the performance of mixed models?

[Gabriel Baud-Bovy]

Another idea might be to use parallellization for bootsrap.
Best,
Gabriel

On 4/18/2013 12:40 AM, Titus von der Malsburg wrote:
>
> Packages like parallel allow you to create temporary copies the current
> R process such that the various copies can work on different problems
> (using differnt cores).  Technically, that's similar to running several
> instances of R in parallel, just more convenient because R automatically
> collects the results of the parallel computations in the parent process.
>
> As you suspected, the execution of a single function like lmer can't be
> split up with these packages.  Automatic parallelization of code written
> for a single execution thread is a very hard problem.  In many cases
> it's even impossible.  Therefore, in order to make use of several cores,
> lmer would have to be rewritten in some way, which may also not be
> trivial.  So, I'm afraid it's currently not possible to use several
> cores and it's very likely going to stay that way for at least some
> time.
>
> For now, I can offer only two ideas: 1.) If you have to run several
> models, e.g. for different dependent variables, you can use parallel to
> fit these models concurrently, each on one core.  2.) If your models are
> fit on very large data sets, it may happen that your computer is running
> out of RAM.  In this case, the operating system will extend the RAM
> using disk space.  The problem with that is that hard disks are several
> orders of magnitude slower than RAM and therefore everything will slow
> down to a crawl.  The solution is then to extend the RAM of your
> computer.  RAM is cheap but you have to make sure that you have a 64 bit
> operating system.  32 bit operating systems can't make use of RAM
> capacities larger than 4 GB.  There is probably an upper limit to how
> much RAM R can use but I don't know that from the top of my head.  If
> your problem really is RAM, then extending RAM will give you a
> tremendous speed-up.
>
> Good luck!
>
>    Titus
>
> Felipe Vargas Reeve writes:
>>     Hi everyone, I want to know if somebody can help me with this: A
>> methodology to increase the performance of R to achieve the convergence in
>> lmer or nlme moldels.
>>
>>     Actually even if the computer presents more than one core I have read
>> that R works with only one of them. Also I have read about the existence of
>> packages (Eg. parallel) that can improve the speed of the computer. This is
>> based in using all the cores of the pc, but I think this works for
>> independent functions and it does not optimize only one process like in the
>> case of the linear mixed model.
>>
>> Thanks for your help guys,
>>
>> Regards
>> Felipe.
>>
>> 	[[alternative HTML version deleted]]
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>


-- 
---------------------------------------------------------------------
Gabriel Baud-Bovy               tel.: (+39) 02 2643 4839 (office)
UHSR University                       (+39) 02 2643 3429 (laboratory)
via Olgettina, 58                     (+39) 02 2643 4891 (secretary)
20132 Milan, Italy               fax: (+39) 02 2643 4892