Quest for a High Performance Boltzmann Transport Solver.

摘要

A 3-d Boltzmann transport code developed at Lawrence Livermore National Laboratory exploits concurrency (problem permitting) with respect to all phase space variables represented by direction, position and energy. To demonstrate the capabilities of the code, a highly resolved Boltzmann transport calculation with 27.6 billion unknowns (332 billion discretization points) was performed on 4992 processors of the IBM RS/6000 SP ASCI Blue machine at Lawrence Livermore National Laboratory. Detailed performance measurements were collected during this calculation. Analyzing the timing data revealed a relatively modest performance, i.e. obtained fraction of peak performance. A detailed study of the performance data confirmed that this modest performance was caused by low utilization of local processor performance, as well as low parallel efficiency of the intrinsically sequential sweep procedure. Both issues and their ramifications will be addressed in this paper.