2014SNA + MC 2013 - Joint International Conference on Supercomputing in Nuclear Applications + Monte Carlo
|Number of page(s)||8|
|Section||4. Advanced Parallelism and HPC Strategies: a. Deterministic Methods, Parallelism and HPC|
|Published online||06 June 2014|
Accelerating COBAYA3 on multi-core CPU and GPU systems using PARALUTION
1 Karlsruhe Institute of Technology (KIT), Institute for Neutron Physics and Reactor Technology (INR) Hermann-von-Helmholtz-Platz 1, Geb. 521, 76344 Eggenstein-Leopoldshafen, Germany
2 Uppsala University, Dept. of Information Technology, Div. of Scientific Computing Lägerhyddsvägen 2, 752 37 Uppsala, Sweden
* Corresponding Author, Email: Nico Trost, firstname.lastname@example.org
COBAYA3 is a multi-physics system of codes which includes two 3D multi-group neutron diffusion codes, ANDES and COBAYA3-PBP, coupled with COBRA-TF, COBRA-IIIc and SUBCHANFLOW sub-channel thermal-hydraulic codes, for the simulation of LWR core transients. The 3D multi-group neutron diffusion equations are expressed in terms of a sparse linear system which can be solved using different iterative Krylov subspace solvers. The mathematical SPARSKIT library has been used for these purposes as it implements among others, external GMRES, PGMRES and BiCGStab solvers.
Multi-core CPUs and graphical processing units (GPUs) provide high performance capabilities which are able to accelerate many scientific computations. To take advantage of these new hardware features in daily use computer codes, the integration of the PARALUTION library to solve sparse systems of linear equations is a good choice. It features several types of iterative solvers and preconditioners which can run on both multi-core CPUs and GPU devices without any modification from the interface point of view. This feature is due to the great portability obtained by the modular and flexible design of the library.
By exploring this technology, namely the implementation of the PARALUTION library in COBAYA3, we are able to decrease the solution time of the sparse linear systems by a factor 5.15x on GPU and 2.56x on multi-core CPU using standard hardware. These obtained speedup factors in addition to the implementation details are discussed in this paper.
Key words: COBAYA3 / diffusion approximation / PARALUTION / acceleration / parallelization / multi-core / GPUs
© Owned by the authors, published by EDP Sciences, 2014