Folding applications into high dimensional torus networks
Arnold Lukas · Advances in parallel computing · 2012
Scientific HPC applications with next neighbour communication pattern can only benefit from torus networks, if the MPI ranks are properly arranged. Otherwise communication time penalties have to be taken into account, as the messaging distance lowers the network's performance. The impact of an individual mapping is shown in synthetic and application benchmarks, for the plasma physical codes PEPC, PSC, and RACOON. A general approach to rank mapping on high dimensional torus network is provided. This approach is based on the combination of spare network dimensions via space filling curves.