I'd like to announce the availability of the new "doMPI" package, a parallel backend for the "foreach" package, which acts as an adaptor to the "Rmpi" package. The package has been uploaded to CRAN and is now available. I wrote this package after leaving REvolution Computing because I liked the foreach package, wanted to learn more about the Rmpi package, and thought it might be of value to the R community by combining the speed of Rmpi with the ease of use of foreach. Perhaps that's an odd way to spend my time when my wife probably thinks I should be looking for a job. I guess I'm obsessive. And I thought it would only take a week or two. I'm not really sure if I succeeded in making a package that is substantially better than the doSNOW package that I wrote previously, at least on smaller clusters where broadcasting isn't very helpful. I ended up adding tuning arguments that allow broadcasting to be turned off, force the use of piggy- backing shared data with the first task, etc. At any rate, I learned a lot, had fun, and gained an even deeper appreciation of the excellent snow package during my testing and benchmarking. I would like to thank Dirk Eddelbuettel and Bryan Lewis for their help on the doMPI package. Both found problems in the package, and gave me valuable feedback. I was also greatly helped by Dirk's excellent "HPC with R" tutorial. It has information on using Rmpi that I didn't find anywhere else. Bryan researched various MPI topics for me, and gave me access to an excellent little cluster for testing purposes. Also note that the package includes a vignette which has information on running doMPI scripts using the orterun command. I'm tempted to write a similar article for snow, since there doesn't seem to be much information available on that subject outside of Dirk's tutorial. - Steve Weston
new doMPI package available on CRAN
1 message · Stephen Weston