[R-sig-ME] MCMCglmm computation time

Jarrod Hadfield j.hadfield at ed.ac.uk
Sun Aug 14 11:38:28 CEST 2011


Dear Stephane,

I should probably allow the option of writing-to-file during an MCMC  
run for those with temperamental power supplies ....  You can merge  
MCMC chains from multiple runs, although you should make sure you  
start them from different initial values and that they pass  
convergence diagnostics. mcmc.list from the library coda is useful for  
manipulating parallel chains.

Cheers,

Jarrod






On 4 Aug 2011, at 14:14, Stephane Chantepie wrote:

> Dear all,
>
> I am a phd candidate working on aging in some bird species. To  
> achieve this,
> I have started to use MCMCglmm a few month ago. To detect  
> senescence, I need
> using large data sets and hence I have to wait about two weeks or  
> often more
> for the runs to complete. My major problem is that the capricious
> electricity network often shut down my computers, so I am looking  
> for ideas
> to reduce computation time (and finish runs).
>
> I have read on the internet that it is too complicated (or  
> impossible) to
> gain time on the algorithm. So, I have decided to use the package  
> "doMC" in
> order to divide the markov chain between the cores of my computers.  
> With
> this, instead of running a model with 4300000 iterations (with 300000
> burning iterations), I would run in parallel four models (one by  
> core) with
> 1300000 iterations (300000 burning iterations and the same prior in  
> the 4
> models).
>
> My question : Is it possible to concatenate results after burning  
> (Sol, VCV,
> deviance) and use it like a single big model or is it conceptually  
> wrong? I
> know that I loose computation time by multiplying burning but is it  
> wrong to
> do this alternative strategy?
>
> Thanks a lot in advance for your input,
>
> Best
>
>
> Stephane
>
> --
> Stephane Chantepie
> CNRS Phd  candidate
> Muséum national d'Histoire naturelle
> 55 rue Buffon
> 75005 paris
> Tel : 01 40 79 32 03
> E-mail : chantepie at mnhn.fr
> --
>
> 	[[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.




More information about the R-sig-mixed-models mailing list