Wilson-Dirac Operator Revisited on Multicore Vector Computers
Résumé
We revisit the Wilson-Dirac operator, also refered as Dslash, on multicore vector machines. The Wilson-Dirac operator is the major computing kernel in Lattice Quantum ChromoDynamics (LQCD), which is the canonical discrete formalism for Quantum ChromoDynamics (QCD) investigations. QCD is the theory of sub-nuclear particles physics, aiming at modeling the strong nuclear force, which is responsible for the interactions of nuclear particles. Based on LQCD formalism, intensive simulations are performed following the Monte Carlo paradigm. Informative observations are expected from large-scale and numerically sensitive LQCD simulations. The corresponding computing demand is therefore tremendous, thus the serious consideration for powerful supercomputers. Designing efficient LQCD codes on modern (mostly hybrid) supercomputers requires to efficiently exploit all available levels of parallelism including accelerators. Since the Wilson-Dirac operator is a coarse-grain stencil computation performed on huge volume of data, any performance and scalability related investigation should skillfully address memory accesses and interprocessor communication overheads. In order the lower the latter, an explicit shared memory implementation should be considered at the node level, since this will lead to a less complex data communication graph. This the main focus of the current paper, where we provide, explain, and discuss a multi-threaded vector implementation, whose experimental results in double precision on the recently released INTEL BROADWELL based machine show a competitive absolute efficiency and a good scalability on one of its four NUMA nodes. An extension to all available nodes is currently under investigation through NUMA-awareness consideration. 545 cuba050new.aketitle
Origine : Fichiers produits par l'(les) auteur(s)
Loading...