» Back to overview
Proposing Institution

Centro Fermi, Rom
Project Manager

Dr. Andrea Pallottini
Piazza dei Cavalieri 7
I-56126 Pisa
This project aims at optimizing the AMR code RAMSES-RT, that we have recently coupled with thethermochemistry solver KROME, as a preparation for PRACE application to Tier-0 projects. This code will be used for zoom-in simulations of the high-redshift universe. In particular we will investigate the physics of early galaxy formation and evolution from cosmological scales to molecular cloud formation. With this code it will be possible to efficiently follow the evolution of galaxies: as the gas falls in dark matter halos located in the knots of the cosmic web, molecular hydrogen forms and cools the gas, eventually leading to the generation of first stars. The energyinput from these luminous sources regulates the infall/outflow and the subsequent generations of stars in a feedback cycle.The study of the formation and evolution of galaxies in a cosmological framework requires numerical codes that are able to follow different physical processes, such as gravity, hydrodynamics, radiative transfer and chemistry (via the KROME module). Standard cosmological codes are optimized to solve gravity and hydrodynamics; hence domain splitting routines focus on optimizing the load balance between these two processes only. However, when radiative transfer and chemistry are included, as in our case, the different computational costs must be carefully balanced bya proper domain decomposition. Furthermore, other key physical processes, as star formation, require the creation of particles, which depends on inter-process communications to avoid data race problems. If the computational domain is not evenly split, different processes will have to mutually wait before advancing to the next time-step. This makes the simulation very inefficient in exploiting the available computational resources, especially for large facilities.In this project, we aim at improving the efficiency of RAMSES-RT code on computer clusters with a large number of nodes, by- implement load balance routines that account for different physical processes;- using MPI-OpenMP hybrid schemes to optimize the computational cost;- minimize MPI wait times by optimizing inter-process communications.

Impressum, Conny Wendler