Abstract: |
|
High performance computer architectures are constantly evolving and to use applications in an efficient way, they have to be ported. In this presentation we will discuss how the compressible fluid flow solver HiPSTAR, used to study transition, turbulence and combustion, is adopted to perform on multi GPU node systems. HiPSTAR has been used to perform large scale simulations on TITAN at the Oakridge National Lab as well as Piz Daint at CSCS that were based on 1 x86 CPU + 1 NVIDIA K20x. In recent time compute nodes comprised of several GPUs and CPUs have been introduced that require to rethink load balancing due to the heterogeneous environment. In addition, Power 8 based GPU systems started to become available. In this contribution we will present developments to HiPSTAR required to perform well on both IBM Minsky as well as X86 based systems comprised of 2 CPUs and 4 GPUs. Performance for a turbulence problem will be shown.
|
|