When it is runned a random walk metropolis hasting in a just one cores one seeks convergence and I would like to understand what is the advantage of running dynare parallel? I ask this because I understand convergence depends of initial point of RWMH (computed with mode) and the initial point it is the same in differents parallel chains. It is not clear to me what is the advantage of dynare parallel in improving computing time given parallel chains depend of the same initial point.
Dear Also,
The initial conditions of the chains are different, otherwise it would be pointless to run more than one chain. The initial conditions of the chains are randomly drawn from a multivariate gaussian centered on the the posterior mode estimated with the optimization routine.
Best,
Stéphane.
Thanks for the answer dear Stepan,
Could you please give me some details about the advantages of running in parallel with dynare parallel in relation to running in a just one core?
Parallel execution is very useful if you are running multiple chains instead of one long chain (which is a matter of preference). In this case, you can run the inherently sequential Monte Carlo chains in parallel, saving considerable time. On top of that, one the estimation has finished and various diagnostics and figures need to be computed, these computations will also be executed in parallel.
However, given Matlab’s automatic use of more than one core for some standard computations, it is not the case that non-parallel execution will only rely on one core. Because of this, it is also not advisable to max out the number of cores when using parallel execution, particularly as overhead needs to be managed as well.