To address the performance and programming challenges from the constantly evolving computer architecture in parallel discrete ordinates method, this paper presents our continue effort on developing and improving the parallel 3D discrete ordinates radiation transport code JSNT-S. A hybrid acceleration method for source iteration is presented, which combines the PCR algorithm and the Anderson algorithm for accelerating different energy groups. A patch-abstraction and hybrid MPI+threads implementation is employed for the domain partition algorithm and the sweeping algorithm. Several optimizations on the data-driven sweeping algorithm are presented, such as exploiting parallelism in angular directions, developing optimal priority strategy for scheduling and reducing DAG construction and scheduling overhead by vertex clustering. The correctness is verified using the AMBER procedure and the result of VENUS-3 benchmark is demonstrated. The parallel performance is measured using the model modified from VENUS-2 and the Kobayashi benchmark, the result of which shows reasonable scaling performance up to 76,800 cores.

This content is only available via PDF.
You do not currently have access to this content.