Towards Optimal Performance for Lattice Boltzmann Applications on Terascale Computers

The performance of many CFD applications, like the lattice Boltzmann method (LBM), on popular cluster computers can fall far short of the impressive peak performance numbers. Using a large scale LBM application, we demonstrate the different performance characteristics of modern supercomputers. Classical vector systems (NEC SX8) still combine excellent performance with a well established optimization approach and can break the TFlops application performance barrier at a very low processor count. Although, modern microprocessors offer impressive peak performance numbers and extensive code tuning has been performed, they only achieve 10% or less of the vector performance for the LBM application. Clusters try to fill this gap by massive parallelization but network capabilities can impose severe performance restrictions if the problem size is not arbitrarily scalable. Tailored HPC servers like the SGI Altix series can relieve these restrictions.