PENTRAN: A three-dimensional scalable transport code with complete phase-space decomposition

Three-dimensional SN transport modeling is gaining importance, with a rising need for detailed fluence calculations in geometries with inherent asymmetries. These calculations often cannot be adequately performed due to immense memory and storage requirements. New scalable, parallel discrete ordinates codes that take advantage of distributed memory parallel architectures are required. In general, scalable parallel algorithms minimize serial operations, maximize algorithm and data parallelism, and maximize computation to communication ratio. A leading issue in scalability is the minimization of serial input/output (I/O) in parallel algorithms. All I/O tasks should be handled efficiently by each individual processor. Further, memory-intensive arrays should be partitioned locally, distributed among processors. Regardless of parallel speedup, use of a local, partitioned memory structure is the only way that larger problems can be solved at all. In addition to using local memory allocation, efficient interprocess communicator groups for specific sets of processors should be used, allowing as needed message passing between necessary groups of processors, avoiding wasteful communications. Acceleration methods are needed because, without them, convergence can be difficult in some problems. Complete parallel decomposition of the problem should also be available; recent analysis of the transport source iteration indicates that significant parallel speedups are possible with various decomposition strategies of the phase-space (angle, energy, and spatial dimensions) over a large processor array