|Title||Hierarchical partitioning of unstructured meshes in cardiac electrophysiology|
|Project(s)||Meeting Exascale Computing with Source-to-Source Compilers, Center for Biomedical Computing (SFF)|
|Publication Type||Talks, invited|
|Year of Publication||2016|
|Location of Talk||Third Workshop on Programming Abstractions for Data Locality (PADAL'16), Kobe, Japan|
Unstructured meshes are widely used in computational science and provide numerous advantages over structured meshes in many applications. W.r.t. data locality however, their irregular data structures and access patterns pose severe challenges, especially on modern heterogeneous clusters with deep memory hierarchies which are poised to become the standard in the foreseeable future. We discuss several of the challenges involved in hierarchical partitioning, such as preserving locality in complex accelerator-equipped nodes, load balancing between heterogeneous devices, and reordering for cache efficiency, as well as solutions to these problems implemented in a cell-centered finite volume simulation code from cardiac electrophysiology. Finally, we present our current steps on the way to a true hierarchically partitioning software for automatic data placement across heterogeneous clusters with varying node architectures.