In this study, a representative benchmark case for the IAG in-house DNS code NS3D is presented, which has been developed specifically with the aim of evaluating code performance on Hawk and future machines on up to full machine size. The case is representative for simulations of compressible turbulent flat plate boundary layers, which are intended to provide both reference data and a better understanding of the physical processes within realistic boundary layers. Weak and strong scaling behaviour is shown up to 4096 nodes, demonstrating good scaling performance. Additionally, NS3D’s newly developed internal performance monitoring tools are presented and the statistics of the computational performance of several important routines are analysed. The performance monitoring gives indications of potential problems during the simulation runtime, such as nodes that are too slow, and can be configured to abort the simulation in case of unacceptable performance. Finally, for the weak scaling cases, the scaling performance of the recently implemented parallel HDF5 I/O is shown.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Computational Aspects for DNS of Compressible Turbulent Boundary Layers on up to Full Machine Size

  • Tobias Gibis,
  • Jason Appelbaum,
  • Christoph Wenzel

摘要

In this study, a representative benchmark case for the IAG in-house DNS code NS3D is presented, which has been developed specifically with the aim of evaluating code performance on Hawk and future machines on up to full machine size. The case is representative for simulations of compressible turbulent flat plate boundary layers, which are intended to provide both reference data and a better understanding of the physical processes within realistic boundary layers. Weak and strong scaling behaviour is shown up to 4096 nodes, demonstrating good scaling performance. Additionally, NS3D’s newly developed internal performance monitoring tools are presented and the statistics of the computational performance of several important routines are analysed. The performance monitoring gives indications of potential problems during the simulation runtime, such as nodes that are too slow, and can be configured to abort the simulation in case of unacceptable performance. Finally, for the weak scaling cases, the scaling performance of the recently implemented parallel HDF5 I/O is shown.