Performance of a neutron transport code with full phase space decomposition on the Cray Research T3D

We present performance results obtained on a 128-node Cray Research T3D computer by a neutron transport code implementing a standard mtiltigroup, discrete ordinates algorithm on a three-dimensional Cartesian grid. After summarizing the implementation strategy used to obtain a full decomposition of phase space (i.e., simultaneous parallelization of the neutron energy, directional and spatial variables), we investigate the scalability of the fundamental source iteration step with respect to each phase space variable. We also describe enhancements that have enabled performance rates approaching 10 gigaflops on the full 128-node machine