Forward modeling of gravitational fields on hybrid multi-threaded cluster
The analytic solution of the gravimetric tensorcomponents, making use of the gravitationalpotential equation for a three-dimensionalvolumetric assembly composed of unit prismsof constant density, demands a high compu-tational cost. This is due to the gravitationalpotential of each one of these prism...
| Autores: | , , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2015 |
| País: | México |
| Institución: | Universidad Nacional Autónoma de México |
| Repositorio: | Redalyc-UNAM |
| OAI Identifier: | oai:redalyc.org:56832992003 |
| Acceso en línea: | https://www.redalyc.org/articulo.oa?id=56832992003 |
| Access Level: | acceso abierto |
| Palabra clave: | Ciencias de la Tierra MPI hyper OpenMP gravity clusters |
| Sumario: | The analytic solution of the gravimetric tensorcomponents, making use of the gravitationalpotential equation for a three-dimensionalvolumetric assembly composed of unit prismsof constant density, demands a high compu-tational cost. This is due to the gravitationalpotential of each one of these prisms must becalculated for all of the points of a previouslydefined observation grid, which turns out in alarge scale computational cost. In this workwe introduce a hybrid design and its parallelimplementation, based on OpenMP and MPI,for the calculation of the vectorial componentsof the gravimetric field and the componentsof the gravimetric tensor. Since the comput-ing time is drastically reduced, the obtainedperformance leads close to optimal speed-upratios. The applied parallelization techniqueconsists of decomposing the problem intogroups of prisms and using different memoryallocations per processing core to avoid bottle-neck issues when accessing the main memoryin one cluster node, which are generally pro-duced when using too many execution threadsover the same region in OpenMP. Due OpenMPcan be only used on shared memory systemsis necessary to use MPI for the calculationdistribution among cluster nodes, giving as aresult a hybrid code (OpenMP+MPI) highly ef-ficient and with a nearly perfect speed-up. Ad-ditionally the numerical results were validat-ed with respect to its sequential counterpart. |
|---|