Scalable Hybrid Parallelization Strategies for the DUNE Grid Interface

Scalable Hybrid Parallelization Strategies for the DUNE Grid Interface
复制标题

DUNE 网格接口的可扩展混合并行化策略

DOI:
--
复制
发表时间:
2013
期刊:
European Conference on Numerical Mathematics and Advanced Applications
影响因子:
--
通讯作者:
J. Fahlke
J. Fahlke
中科院分区:
--
文献类型:
--
作者:
C. Engwer;J. Fahlke

文献摘要

被引文献

相似文献

DUNE框架提供了一个既灵活又高效的PDE工具箱。为了保持现代和未来架构的性能,必须将面向硬件的技术集成到 DUNE 中。我们介绍了当前向 DUNE 网格接口添加混合并行化的工作,该接口迄今为止仅支持 MPI 并行化。在当前的硬件趋势中,我们看到了从多核架构到众核架构的转变,例如英特尔 PHI。在传统多核 CPU 上运行良好的技术在多核系统上不再适用。我们比较了向 DUNE 添加线程并行层的不同策略,并讨论了它们的可扩展性和性能。
The DUNE framework provides a PDE toolbox which is both flexible and efficient. Integration of hardware oriented techniques into DUNE will be necessary to maintain performance on modern and future architectures. We present the current effort to add hybrid parallelization to the DUNE grid interface, which up to now only supports MPI parallelization. In current hardware trends, we see a transition from multi-core to many-core architectures, like the Intel PHI. Techniques which worked well on traditional multi-core CPUs don’t scale anymore on many-core systems. We compare different strategies to add a thread parallel layer to DUNE and discuss their scalability and performance.