Background: I am estimating a large model using mode_compute = 6, but it is taking ages (and ages) to run so I am investigating options for parallel computing. I am at a university in South Africa (Stellenbosch University) and we have a High Performance Computer (HPC) cluster setup that can be accessed remotely using Matlab’s Parallel Computing Toolbox (PCT).
Mode computing does not rely on parallelization as is. So going parallel would not help in that step.
mode_compute=6 is a truly inefficient mode-finder. Is there a reason you need it? Most of the times, people use it to hide other issues in their model.
Harri, I agree with Johannes. mode_compute=6 is really only useful if your struggling to get a non-positive definite mode. If it finds a mode for you, then input that mode using mode_file= with mode_compute=4. Set mh_replic=100000, mh_nblocks=5, mh_drop=0.5 and see if the algorithm can get parameter estimate convergence (the diagnostic statistics). If you’re not able to even get the estimation going (ie, mode_compute=4 gives you non-positive definite mode, even after inputting the mode from mode_compute=6… there’s defintiely an issue with the model.)
Happy to go through the code with you sometime… just pop me an email.