Dynamic Load Balancing for Parallel Mesh Adaptation

Computational fluid dynamics (CFD) flow simulations are extremely expensive in terms of CPU time and memory. In this study, parallel computing and grid adaptation techniques are employed to achieve high efficiency and accuracy in a hybrid unstructured flow solver. However, adaptive local grid refinement/coarsening causes the unequal distribution of workload among the processors at run time. A simple, effective repartitionand remapping-based dynamic load balancing scheme, named RARB, has been developed and integrated into the flow solver to solve the load imbalance problem. As first major component of RARB, a modified Recursive Coordinate Bisection (RCB) partition algorithm is exploited to repartition the computational domain due to its simplicity and efficiency once the load imbalance is detected. Two heuristic rules have been used to facilitate remapping the new partitioned sub-domains to the processors with less data communication cost. Task migration from overloaded processors to underloaded processors is the second major component of RARB and is handled in parallel by a multi-level granularity procedure. In addition, three metrics have been used as an indicator of the global view of system load. The Experiments conducted on a cluster of PCs show high efficiency and accuracy of the flow solver to accomplish complex flow computations, and the effectiveness of RARB to handle the load imbalance in grid adaptations. Key-Words: Computational fluid dynamics, Dynamic load balancing, Solution adaptation, Mesh partition, Parallel and distributed application, Flow solver, Parallel algorithms