(Paper #32)
This paper discusses parallel scaling performance of large scale parallel structured adaptive mesh refinement (SAMR) calculations in SAMRAI. Previous work revealed poor scaling in certain adaptive gridding operations in SAMR calculations. These operations are of minimal cost on small numbers of processors but become dominant for large-scale problems run on up to 512 processors. This work describes algorithms we have developed to enhance the efficiency of the adaptive gridding operations. Performance of the algorithms is evaluated for two adaptive benchmarks run on up 512 processors of an IBM SP system.
Keywords:
Scientific Computing
Profiling, Instrumentation and Measurement
Numerical Algorithms
Case Study/Experience Report
Performance Modeling and Evaluation