Computational strategies for implementation of 2D elastic wave modeling in GPU
Elastic wave modeling presents a challenge to implement since it is a computationally costly procedure. Nowadays, due to GPU increased power jointly with development in HPC computation, it is possible to execute elastic modeling with better execution times and memory use. This study evaluates the pe...
- Autores:
- Tipo de recurso:
- Article of journal
- Fecha de publicación:
- 2020
- Institución:
- Universidad Católica de Pereira
- Repositorio:
- Repositorio Institucional - RIBUC
- Idioma:
- spa
- OAI Identifier:
- oai:repositorio.ucp.edu.co:10785/10033
- Acceso en línea:
- https://revistas.ucp.edu.co/index.php/entrecienciaeingenieria/article/view/2016
http://hdl.handle.net/10785/10033
- Palabra clave:
- Rights
- openAccess
- License
- Derechos de autor 2021 Entre Ciencia e Ingeniería
| Summary: | Elastic wave modeling presents a challenge to implement since it is a computationally costly procedure. Nowadays, due to GPU increased power jointly with development in HPC computation, it is possible to execute elastic modeling with better execution times and memory use. This study evaluates the performance of 2 strategies for implementing elastic modeling using different kernel launching layouts, CPML memory allocation strategies, and wavefield storage management. The performance measures show that the algorithm, which includes 2D kernel launching layout, CPML reduced memory strategy, and GPU global memory storage to save wavefield cube peaks up to 88.4% better execution time and uses 13.3 times less memory to obtain the same elastic modeling results. There is also an increasing trend of enhancement in execution times and memory savings when working with models of bigger sizes with this strategy. |
|---|