Running time

June 10, 2013
By

(This article was first published on Gianluca Baio's blog, and kindly contributed to R-bloggers)

Marta and I are doing some re-analysis of our Eurovision contest (some context here and here). We have slightly modified our original model (mostly, I have navigated the mess in Marta's notation $-$ it's OK: I'm not at risk of her mighty wrath, as I've already joked about this with her!). Also, we've included the latest data (in fact, we're doing a little prediction as well, pretending that all the covariates are known, but the actual results aren't).

Last time around, Marta ran the model using parallelisation (through the R package snowfall) in WinBUGS on her Windows desktop, at work. The model is relatively complex (see here for a working description), but thanks to parallelisation the running time was not too bad (as I recall, about 3 hours to run 11000 iterations with a burn-in of 1000 and thinning of 20).

This time, I decided I would get my hands dirty on my Linux machine(s). To start with, I thought I'd give it a go using R2jags and its new(-ish) function jags.parallel, which is supposed to run multiple instances of the MCMC process at once. Unfortunately, that didn't work. I searched on line for some clue and found people sharing their common experiences. I don't really know what's failing, but JAGS complains that a node (which I'm actually passing as data) is not found.

So, thinking that running software natively rather than via emulation would be most efficient, I decided to install OpenBUGS and run the model using snowfall and R2OpenBUGS. This was better: I was able to check the model and load the data. But when it came to compiling the model, OpenBUGS took for ever $-$ I think for some reasons it was getting stuck with something. It was a bit surprising, as I thought that this would be a very robust and reliable alternative.

So, I decided to try and see how WinBUGS would do. I installed WINE and ran the model using R2WinBUGS and again through snowfall $-$ effectively replicating the exact procedure that Marta used last year. To my surprise, things have worked out quite OK and I shaved about an hour from Marta's previous attempt (mainly through blocking of some of the parameters). 

To leave a comment for the author, please follow the link and comment on his blog: Gianluca Baio's blog.

R-bloggers.com offers daily e-mail updates about R news and tutorials on topics such as: visualization (ggplot2, Boxplots, maps, animation), programming (RStudio, Sweave, LaTeX, SQL, Eclipse, git, hadoop, Web Scraping) statistics (regression, PCA, time series, trading) and more...



If you got this far, why not subscribe for updates from the site? Choose your flavor: e-mail, twitter, RSS, or facebook...

Comments are closed.