Parallel streamline tracking for UnTRIM

Jacek A. Jankowski. Parallel streamline tracking for UnTRIM, 4th UnTRIM Workshop, Trento, 7-9 May 2007.

The parallel streamline tracking algorithm for the domain decomposition method applied in the MPI version of UnTRIM, is described in detail. The streamline tracking is the integral part of the semi-Lagrangian method of characteristics, being one of the most robust methods for the non-linear momentum advection, working also for larger Courant numbers. The method is however awkward to parallelise in the framework of the point-to-point communication typical for the domain decomposition method.

An autonomous algorithm is applied for treatment of tracebacks leaving the partition they stem from. Treated as MPI objects, they are followed through neigbour partitions. The final results of value interpolation at the end of tracebacks are delivered by global communication back to the partitions of the traceback origin. The careful verification revealed some small result reproducibility issues for different number of partitions due to the slight mesh (sorting) dependency of the original serial algorithm.

The scalability of the developed parallel algorithm is proven by comparisons between runs with and without the momentum advection stage. The overall scalability of the code is not adversely affected due to the fact that only tracebacks leaving a partition and its halo by higher local Courant numbers are considered. The scalability of the complete MPI UnTRIM algorithm diminishes only when the number of halo cells exceeds ca. 10%.

Transparencies are available.

See also the journal paper in the Int. J. Num. Meth. in Fluids.