Parallel Computational Fluid Dynamics: Not without its Challenges

摘要

The challenges of parallelisation for CFD have long been addressed. The strategy of load balancing through mesh partitioning is now well established and there are many tools around to support this strategy. However, although these strategies are well established and work really well in some contexts, scalability is not guaranteed for all applications on all systems. Many CFD codes do not yield the scalability that might have been expected from the results reported in the early to mid 1990's. This paper explores some of the reasons for the limited scalability, which can arise from a series of causes: a) The hardware - especially in relation to the ratio of the processor compute speed to the combined latency and bandwidth of the inter-processor communications b) The nature of some problems which are naturally heterogeneous in the compute load across the mesh (such as combustion fronts or free surfaces) c) Differential mesh movement (such as for locally rotating meshes) The exploration will be in the context of a finite volume code using a heterogeneous unstructured mesh which has been parallelised using conventional mesh partitioning strategies and an effective load balancing tool, JOSTLE.