Parallelizing XDEM: Load-balancing policies and efficiency, a study

In XDEM, the simulation domain is geometrically decomposed in regular fixed-size cells that are
used to distribute the workload between the processes. The role of the partitioning algorithm is to
distribute the cells among all the processes in order to balance the workload. To accomplish this task,
the partitioning algorithm relies on a computing/communication cost that has been estimated for each
cell. A proper estimation of these costs is fundamental to obtain pertinent results during this phase.
The study in the work is twofold. First, we integrate five partitioning algorithms (ORB, RCB, RIB, kway
and PhG) in the XDEM framework [1]. Most of these algorithms are implemented within the
Zoltan library [2], a parallel framework for partitioning and ordering problems. Secondly, we propose
different policies to estimate the computing cost and communication cost of the different cells
composing the simulation domain. Then, we present an experimental evaluation and a performance
comparison of these partitioning algorithms and cost-estimation policies on a large scale parallel
execution of XDEM running on the HPC platform of the University of Luxembourg. Finally, after
explaining the pros and cons of each partitioning algorithms and cost-estimation policies, we discuss
on the best choices to adopt depending on the simulation case.