Performance of a Block Structured, Hierarchical Adaptive MeshRefinement Code on the 64k Node IBM BlueGene/L Computer
We describe the performance of the block-structured Adaptive Mesh Refinement (AMR) code Raptor on the 32k node IBM BlueGene/L computer. This machine represents a significant step forward towards petascale computing. As such, it presents Raptor with many challenges for utilizing the hardware efficiently. In terms of performance, Raptor shows excellent weak and strong scaling when running in single level mode (no adaptivity). Hardware performance monitors show Raptor achieves an aggregate performance of 3:0 Tflops in the main integration kernel on the 32k system. Results from preliminary AMR runs on a prototype astrophysical problem demonstrate the efficiency of the current software when running at large scale. The BG/L system is enabling a physics problem to be considered that represents a factor of 64 increase in overall size compared to the largest ones of this type computed to date. Finally, we provide a description of the development work currently underway to address our inefficiencies.
- Research Organization:
- Lawrence Berkeley National Lab. (LBNL), Berkeley, CA (United States)
- Sponsoring Organization:
- USDOE. Office of Advanced Scientific Computing Research.Mathematical Information and Computational Sciences Division, NationalNuclear Security Administration, Lawrence Livermore NationalLaboratory
- DOE Contract Number:
- DE-AC02-05CH11231
- OSTI ID:
- 861020
- Report Number(s):
- LBNL-57500; R&D Project: K11119; BnR: DP1511000; TRN: US200601%%586
- Resource Relation:
- Conference: SC|05 International Conference for HighPerformance Computing Networking and Storage, Seattle, WA, November12-18, 2005
- Country of Publication:
- United States
- Language:
- English
Similar Records
Early Evaluation of IBM BlueGene/P
Scientific Application Performance on Candidate PetaScalePlatforms