Abstract
The present paper studies solvers for partial differential equations that work on dynamically adaptive grids stemming from spacetrees. Due to the underlying tree formalism, such grids efficiently can be decomposed into connected grid regions (clusters) on-the-fly. A graph on those clusters classified according to their grid invariancy, workload, multi-core affinity, and further meta data represents the inter-cluster communication. While stationary clusters already can be handled more efficiently than their dynamic counterparts, we propose to treat them as atomic grid entities and introduce a skip mechanism that allows the grid traversal to omit those regions completely. The communication graph ensures that the cluster data nevertheless are kept consistent, and several shared memory parallelization strategies are feasible. A hyperbolic benchmark that has to remesh selected mesh regions iteratively to preserve conforming tessellations acts as benchmark for the present work. We discuss runtime improvements resulting from the skip mechanism and the implications on shared memory performance and load balancing. © 2013 Springer-Verlag.
| Original language | English (US) |
|---|---|
| Title of host publication | Lecture Notes in Computer Science |
| Publisher | Springer Nature |
| Pages | 484-496 |
| Number of pages | 13 |
| ISBN (Print) | 9783642400469 |
| DOIs | |
| State | Published - 2013 |
| Externally published | Yes |
Bibliographical note
KAUST Repository Item: Exported on 2020-10-01Acknowledged KAUST grant number(s): UK-c0020
Acknowledgements: This work was supported by the German Research Foun-dation (DFG) as part of the Transregional Collaborative Research Centre “Inva-sive Computing (SFB/TR 89). It is partially based on work supported by AwardNo. UK-c0020, made by the King Abdullah University of Science and Technology(KAUST). All software is freely available athttp://www5.in.tum.de/sierpinski.
This publication acknowledges KAUST support, but has no KAUST affiliated authors.