Shared Memory Parallelism for 3D Cartesian Discrete Ordinates Solver

Abstract

This paper describes the design and the performance of DOMINO, a 3D Cartesian SN solver that implements two nested levels of parallelism (multicore+SIMD) on shared memory computation nodes. DOMINO is written in C++, a multi-paradigm programming language that enables the use of powerful and generic parallel programming tools such as Intel TBB and Eigen. These two libraries allow us to combine multi-thread parallelism with vector operations in an efficient and yet portable way. As a result, DOMINO can exploit the full power of modern multi-core processors and is able to tackle very large simulations, that usually require large HPC clusters, using a single computing node. For example, DOMINO solves a 3D full core PWR eigenvalue problem involving 26 energy groups, 288 angular directions (S16), 46 × 106 spatial cells and 1 × 1012 DoFs within 11 hours on a single 32-core SMP node. This represents a sustained performance of 235 GFlops and 40:74% of the SMP node peak performance for the DOMINO sweep implementation. The very high Flops/Watt ratio of DOMINO makes it a very interesting building block for a future many-nodes nuclear simulation tool.

3D Cartesian Transport Sweep for Massively Parallel Architectures with PaRSEC

“…This last change can divide by up to four the arithmetic intensity of the kernel. This confirms the preliminary results on shared memory systems with INTEL TBB presented in [18] against the DENOVO and PENTRAN code.…”

“…However, given that modern supercomputer architectures are becoming more and more heterogeneous (presence of accelerators inside computing nodes) and hybrid (interconnection of several nodes), it may be important to review classical parallel programming models as shown in the paper [17]. In a previous work [18], we have presented the DOMINO neutron transport solver designed for those modern architectures. We have especially showed that: 1) a good data locality dramatically improves arithmetic intensity of the sweep operation, and allows us to efficiently exploit SIMD units available inside current processors; 2) usage of the task-based programming model helped us to parallelize the sweep of DOMINO, by relying on INTEL TBB [19] library that addresses shared memory supercomputing nodes.…”

“…Here only one kernel is required to update angular fluxes in one quadrant (or octant in 3D) for one cell. It is vectorized over angular directions thanks to the C++ generic library Eigen [27] and has been presented in [18].…”

An improved parallelism scheme for deterministic discrete ordinates transport

“…The performance of a shared memory node has also been investigated [17]. Intel Thread Building Block (TBB) tasks were used to maintain a task dependency graph of cells within a wavefront.…”

Expressing Parallelism on Many-Core for Deterministic Discrete Ordinates Transport

“…The performance of a shared memory node has also been investigated [13]. Intel Thread Building Block (TBB) tasks were used to maintain a task dependency graph of cells within a wavefront.…”