Spectral evolution simulation on leading multi-socket, multicore platforms

摘要

Spectral evolution simulations based on the observed Very Long Baseline Interferometry (VLBI) radio-maps are of paramount importance to understand the nature of extragalactic objects in astrophysics. This work analyzes the performance and scaling of a spectral evolution algorithm on three leading multi-socket, multi-core architectures. We evaluate three parallel models with different levels of data-sharing: a sharing approach, a privatizing approach and a hybrid approach. Our experiments show that the data-privatizing model is reasonably efficient on medium scale multi-socket, multi-core systems (up to 48 cores) while regardless algorithmic and scheduling optimizations, sharing approach is unable to reach acceptable scalability on more than one socket. However, the hybrid model with a specific level of data-sharing gives the best scalability over all the considered multi-socket, multi-core systems.