parallel algorithm of OF
 

Hello OFers
I am trying to understand parallelization of OF solvers such as PCG. So far I understood that all processors must follow the same execution path due to blocking calls such as MPI_Allreduce in each iteration. So I wanted to try something else in my solver that avoids synchronization and is more flexible in general. It works by exchanging ghost cell values while the iterations are running. This is not forced but each processor sends back its halo layer values when it recieves one.

The result is that it seems to work well for gauss-siedel/jacobi kind of solvers with small stencils but have problems for CG solvers. I like the flexibiltiy of this approach since two processors can do completely different things (different solvers,fluid-structure etc), as long as they continually exchange information and hopefully convergence is arrived. It seems to me gauss-seidel is ok with this approach since it is kind of 'self-correcting' but a faster algorithm like CG is needed.

Has this parallelization approach been tried in OF before? What kind of parallelization is used in fluid-structure type solutions when two processors are not doing the same thing?

Thanks for any help

I am trying to find if asynchronous CFD calculations are worth a try to include in my thesis, so any tips and directions would be very helpful. I could not find enough material on OF's implementations except a few of Prof. Jasaak slides. Some other papers mention that asynchronous algorithms are fault tolerant and can scale to thousands of processors, but convergence properties are not well defined.

Thanks

mark......