MIGAL and library case E102
 

Hello!

I am looking for anyone who has any experience with MIGAL, preferably for transient, geophysical, stratified BFC-applications. I recently did some trial runs with MIGAL (PHOENICS 3.4), running through some of the examples (E001-E008) and tutorials (E100-E103) in the ecyclopaedia entry for MIGAL.

To my surprise E102, as well as some of the other tutorial runs, ran slower when using MIGAL, compared to using SIMPLEST. Is this to be expected? When is MIGAL a good choice and when is it not?

Thanks,

Olof

Re: MIGAL and library case E102
 

Dear Olof,

Each sweep with MIGAL is 5 to 10 time more expensive (CPU Time) than a SIMPLEST sweep. But the number of sweeps needed to reach a given residual level is much lower with MIGAL. Typically 20-30 sweeps for laminar flows and 60-80 sweeps for turbulent flows. The second point is that these numbers of sweeps are almost grid independent with MIGAL when they increase linearly with the number of cells with SIMPLEST. Hence the finest is your grid the better will be the speed up.

The tutorial E102 just exemplify the use of MIGAL. It is not a speed-up demo case.

I would like also encourage you to test the new MIGAL version delivered with PHOENICS 3.5. It is much more robut and provide some new features like a GMRES multi-grid peconditionned coulped solver. The PHOENICS users conference paper related to this new MIGAL version is downloadable from our web site: www.mfrdc.com

Best regards,

Michel Ferry

Re: MIGAL and library case E102
 

As a side, Is MIGAL still an add on? Advice on how to get it might be good.

Through CHAM or direct to MFRDC?

Thanks in advance.

Re: MIGAL and library case E102
 

The PHOENICS implementation of MIGAL is still a PHOENICS option provided by CHAM and CHAM's local agents.

Best regards,

Michel Ferry

Re: MIGAL and library case E102
 

Dear Michel,

thanks for the explanation. I've downloaded your paper and will take a look at it.

As for my problem, me and my colleagues tend to be a little careless about achieving convergence within each time step. We usually get reasonable results for transient runs with only about 10 sweeps per time step, and neither sweeps nor iterations are stopped by reaching sufficiently low residuals. This inaccurate approach is justifiable, we think, as there are many other error due to inaccurate physics (when considering geophysical, natural flows) and the fact that we otherwise would be unable to perform the calculations within an acceptable time frame. We are already close to a one-to-one relation between simulated time and cpu-time...

Regards,

Olof