High-performance P3M N-body code: CUBEP3M

Abstract

This paper presents CUBEP3M, a publicly-available high performance\ncosmological N-body code and describes many utilities and extensions that have\nbeen added to the standard package. These include a memory-light runtime SO\nhalo finder, a non-Gaussian initial conditions generator, and a system of\nunique particle identification. CUBEP3M is fast, its accuracy is tuneable to\noptimize speed or memory, and has been run on more than 27,000 cores, achieving\nwithin a factor of two of ideal weak scaling even at this problem size. The\ncode can be run in an extra-lean mode where the peak memory imprint for large\nruns is as low as 37 bytes per particles, which is almost two times leaner than\nother widely used N-body codes. However, load imbalances can increase this\nrequirement by a factor of two, such that fast configurations with all the\nutilities enabled and load imbalances factored in require between 70 and 120\nbytes per particles. CUBEP3M is well designed to study large scales\ncosmological systems, where imbalances are not too large and adaptive\ntime-stepping not essential. It has already been used for a broad number of\nscience applications that require either large samples of non-linear\nrealizations or very large dark matter N-body simulations, including\ncosmological reionization, halo formation, baryonic acoustic oscillations, weak\nlensing or non-Gaussian statistics. We discuss the structure, the accuracy,\nknown systematic effects and the scaling performance of the code and its\nutilities, when applicable.\n