Re: [Gretl-devel] gretl-MPI process IDs

On Sat, 14 Jan 2017, Allin Cottrell wrote: > This relates to the thread started By Sven in > http://lists.wfu.edu/pipermail/gretl-devel/2017-January/007254.html > but I think it merits a thread of its own. > > When working up an MPI-based variant of the task Sven's talking about > (launching a bunch of parallel processes, each of which carries out the > same basic computation but with a process-specific set of parameter values) > I noticed something that's a little awkward. Namely, for almost everything > that's indexable in hansl, the index starts at 1, but with MPI we've > carried over the underlying MPI specification in which the indexation of > processes starts at 0 (so n parallel processes have IDs 0 to n-1). > > So long as what's in place is clearly documented (and I think it is), a > hansl coder should be able to handle this without huge difficulty. > Nonetheless I wonder if we'd be better to identify MPI processes in hansl > using a 1-based index. This would be easy to arrange, but would be > backward-incompatible. Which leads to two questions: > > 1) Does anyone at this point have an investment in gretl-MPI scripts that > use 0-based process IDs? (I suspect not, but maybe I'm wrong.) I really don't think so. And even if it were so, anyone who's savvy enough to write scripts using our MPI functions in a non-trivial way should have no problem adjusting to the new convention. > 2) If backward-compatibility turns out not to be a serious issue, do we > think it's worthwhile to switch to 1-based MPI process IDs? Having said the above, I don't see huge advantages either; In fact, the idea that there is a "#0 worker", perhaps makes it a little bit more evident that that worker has a somewhat different, specialised role. However if others feel otherwise, I wouldn't insist.