Tree-SPH code EvoL: profiling and optimization
U. Buonomo · Padua Research Archive (University of Padova) · 2012
This thesis is part of a larger project devoted to modeling galaxies via numerical hydrodynamical simulations, from their formation within their cosmological background throughout their evolutionary history. The thesis has two primary objectives, strictly linked with one another. The first one, of astrophysical nature, is to develop a purely theoretical scenario for the formation and evolution of galaxies in the cosmological context. Given the complexity of the targeted physical systems, it is necessary to rely upon on numerical simulations which in turn require an appropriate numerical code, able to follow the complex physics of galaxies. Therefore, the second objective is the optimization of such code, according to the paradigms of modern computational science. Indeed, given the complexity of the involved physical phenomena the code must be enriched with a number of complex algorithms, which at the same time slow down its performance to such a point that it is difficult to extensively use it to explore the whole scenario in detail, unless sophisticated optimization techniques are applied to drastically reduce the computational time. In this thesis I have studied the problem of the optimization of numerical codes in serial and parallel architectures and I will present the different strategies I have adopted to make them faster, with particular attention to EvoL, the N-body Tree-SPH code entirely written in the house and widely used by the Padua group. To develop and optimize a code like EvoL, specific skills are needed, ranging from classical astrophysics to computer science. To cope with this, in the thesis I first introduce some general concepts on the functioning of a computer machine, its structure and architecture, the various reasons why a code, although correct from a mathematical and logical point a view, does not in practice work at the maximum level of efficiency and does not exploit the formal computing power to disposal in a optimal form. To this aim, I examine some important aspects of the general behaviour of a code and highlight several strategies that can be followed to optimize it. In particular, I analyze the code response in different situations highlighting the limitations, the routines with the heaviest load, and the importance of the communications between CPUs in parallel environment. Given these premises, we present EvoL, a lagrangian code based on the classical NB-TSPH formalism, and describe in some detail its structure and physical content. In particular we report on the new physical treatment of the interstellar medium and the companion codes ROBO and MaNN we have recently developed, and how they are implemented in EvoL. Then we describe the various optimizations we have applied to EvoL in order to improve its efficiency. Finally we report on the physical results we have obtained using EvoL, and ancillary codes ROBO and MaNN. In brief, we describe the physical content of the last release of EvoL, and the many hydrodynamical test we have performed to validate it. Second we describe the two studies dedicated to the treatment of the interstellar medium with the state of the art description of its chemistry and thermodynamical properties, and how these results can be implemented in EvoL, thanks to an extensive use of the artificial neural networks. Finally we present a set new models of early type galaxies obtained from cosmological initial conditions at varying the initial mass and density contrast with respect to the cosmological background and their subsequent evolution into well behaved galaxies. Particular attention is paid to the different role played by the density contrast of dark matter in the proto-halo generating the galaxy and the density of the baryonic matter (gas) required to initiate the star formation process. In general, while proto-galaxies of high mass give origin to a single dominant episode of star formation taking place very early on in a galaxy’s history, proto-galaxies of low mass have a more complicated history of star formation, which a series of episodes of variable intensity often interrupted by periods of quiescence.