Hi, here's a question about some details of random number generation in MPI-parallelized scripts. I have read the relevant section 7 of the gretl + MPI guide, but one thing still isn't clear to me: How to imitate the effect of what "set seed" would do in the standard single-thread setting? As far as I understand, the default setting in gretl's MPI world is to have separate RNGs, to preserve independence of the pseudo-random numbers. OK. The alternative that is described in the guide is to enforce identical RNG sequences in all parallel threads. What about the case where you want to have independent RNG sequences across processes, but you want to be able to control them for replication purposes, from one run to the next? How to achieve that effect in gretl's MPI implementation?

Am 19.07.2020 um 22:04 schrieb Allin Cottrell: > On Sun, 19 Jul 2020, Sven Schreiber wrote: >> What about the case where you want to have independent RNG sequences >> across processes, but you want to be able to control them for >> replication purposes, from one run to the next? How to achieve that >> effect in gretl's MPI implementation? > > set seed > > <hansl> > print "starting MPI..." > mpi >   set seed 88775531 Aha, that's of course very simple. But it feels kind of strange to set just one seed to manage three (or N) independent processes. Do they really stay "independent"?

Am 20.07.2020 um 02:25 schrieb Allin Cottrell: > On Sun, 19 Jul 2020, Sven Schreiber wrote: >> >> Aha, that's of course very simple. But it feels kind of strange to set >> just one seed to manage three (or N) independent processes. Do they >> really stay "independent"? > > They're distinct RNGs under dcmt, so setting the same seed for each > should not induce any dependence. If you're uncomfortable about that you > could set a different seed per MPI process, but if you're willing to > believe that dcmt does what it says it does -- > http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/DC/dc.html -- > that's not necessary. OK, thanks a lot. One final (?) question, namely about the comparison of a single-thread result with a fixed seed to a faster MPI result, also with a fixed seed. I guess it is not possible to make them equal? What I mean is: The naive optimum would be that when I stack the N sequences I get the same set of numbers as I did in the single thread. The purpose would be cross-machine portability and comparability. But again, I expect it isn't easy or possible, but I may be wrong.