NUMA-Aware strategies for the efficient execution of CFD simulations on CPU supercomputers

The growing variety of computing architectures and the hybridization of high-performance computing systems encourage the research for portable implementations of numerical methods in simulation codes. However, the pursuit of efficient and portable implementations is a rather complex problem. The present work is devoted to the development of portable parallel algorithms, primarily for scale-resolving, time-accurate simulations of incompressible flows with turbulent heat and mass transfer. The heterogeneous computing capability allows for engaging both processors and accelerators efficiently. In addition to computing on accelerators, special attention is paid at efficiency on multiprocessor nodes with significant non-uniform memory access factor. In this work, we study in detail the parallel efficiency and performance for different execution modes on up to ten thousand cores of MareNostrum 4 supercomputer.The work has been financially supported by a competitive R+D project (ENE2017-88697-R) by the Spanish Research Agency. X. A. F. is supported by a predoctoral contract (2019FI B2-00076) by the Government of Catalonia. The work has been carried out using the computing resources of the Barcelona Supercomputing Center (MareNostrum 4); The authors thankfully acknowledge the institution.Peer ReviewedPostprint (published version