Optimization of the N-body Simulation on Intel’s Architectures Based on AVX-512 Instruction Set

Abstract

The N-body simulations have become a powerful tool to test the gravitational interaction among particles, ranging from a few bodies to complete galaxies. Even though N-body has already been optimized on many parallel platforms, there are hardly any studies which take advantage of the latest Intel architectures based on AVX-512 instruction set. This SIMD set was initially supported by Intel’s Xeon Phi Knights Landing (KNL) manycore processors launched at 2016. Recently, it has been included in Intel’s general-purpose processors too, starting at the Skylake (SKL) server microarchitecture and now in its successor Cascade Lake (CKL). This paper optimizes the all-pairs N-body simulation on both current Intel platforms supporting AVX-512 extensions: a Xeon Phi KNL node and a server equipped with a dual CKL processor. On the basis of a naive implementation, it is shown how the parallel implementation (can) reach, through different optimization techniques, 2355 and 2449 GFLOPS on the Xeon Phi KNL and the Xeon CKL platforms, respectively.