Large-scale performance analysis of PFLOTRAN with Scalasca
Brian J. N. Wylie, Markus Geimer · JuSER (Forschungszentrum Jülich) · 2011
The PFLOTRAN code for multiphase subsurface flow and reactive transport has featured prominently in US De- partment of Energy SciDAC and INCITE programmes, where is has been used to simulate migration of radionucleide contaminants in groundwater. As part of its ongoing development, execution performance with up to 128k processor cores on Cray XT and IBM BG/P systems has been investigated, and a variety of aspects have been identified to inhibit PFLOTRAN performance at larger scales using the open-source Scalasca toolset. Scalability of Scalasca measurements and analyses themselves, previously demon- strated with a range of applications and benchmarks, required re-engineering in key areas to handle the complexities of PFLOTRAN executions employing MPI within PETSc, LAPACK, BLAS and HDF5 libraries at large scale.