Domain Management for Future Scale Parallelism

One of the main challenges in numerical simulation of real world phenomena on modern high performance clusters is the decomposition and distribution of the overall work load -- especially for growing problem sizes and larger systems with increasing heterogeneity. In this presentation, we introduce our CFD Framework which has been successfully deployed on up to 140,000 cores on two of Germany’s top-tier supercomputing systems. For domain decomposition we rely on non-overlapping block-structured Cartesian Grids gained through hierarchical space-tree refinement, perfectly suited for a custom tailored parallel multigrid-like solver. While domain management is centralised -- using a dedicated management server, aware of the complete topological view of the data structure --, load balancing is realised using space-filling curves. This approach is, however, limited by the available memory of these management servers. As remedy, we propose a novel approach currently under development, where every sub-domain has only a local view of its spatial and hierarchical neighbours. Load balancing is then realised by a diffusion process that will be solved using our multigrid-like approach across the complete hierarchy.     «

One of the main challenges in numerical simulation of real world phenomena on modern high performance clusters is the decomposition and distribution of the overall work load -- especially for growing problem sizes and larger systems with increasing heterogeneity. In this presentation, we introduce our CFD Framework which has been successfully deployed on up to 140,000 cores on two of Germany’s top-tier supercomputing systems. For domain decomposition we rely on non-overlapping block-structured...     »