Parallel 3D Sweep Kernel with PARSEC

The Boltzmann Transport Equation (BTE) is a fundamental law governing the evolution of large sets of particles. BTE numerical solvers are involved in various physical contexts such as nuclear core simulation. A large class of BTE solvers are based on a so-called sweep algorithm that represents the vast majority of the computing operations. In this work, we present an implementation of this sweep algorithm on top of the Parsec runtime. This implementation combines efficiently three levels of parallelism: SIMD, Multi-Threads, and Message Passing. Measured performance on a supercomputer having 768cores reaches 6.2 Tflop/s, and corresponds to 34.4% of the theoretical peak performance of this supercomputer.