rattoma marco.ratto@jrc.it writes:
One minor thing is that we also discussed an option to just run dynare for testing the configuration file, someting like parallel_test.
Thanks for reminding me this, I added it to the wiki page.
Then there is another issue concerning seeds for the MCMC paralle chains. Currently, this is done in metropolis_hastings_initialization.m where the seeds for the parallel chains are generated using the clock. Now I see that in the most recent versions of MATLAB there is a new random number generation engine which allows to properly initiate a number of independent random sequences, which is exactly what we need (RandStream etc.). This would allow, by resetting seeds at the beginning of any dynare execution, to make the chains replicable in a deterministic way. We will have also to test for older matlab versions, where we may leave the clock implementation. What do you think?
This issue is not specific to the parallelization scheme, and is already the object of trac ticket #85: https://www.dynare.org/trac/ticket/85
Best,