Open Access
Subscription Access
On the Performance of the Miranda CFD Code on Multicore Architectures
Abstract
Computational Fluid Dynamics is an important workload on today’s high performance computing systems and has been shown to scale to large numbers of nodes. In this paper we describe our experiences with Miranda, a high order 3D hydrodynamics code for computing fluid instabilities and turbulent mixing. We first demonstrate its scalability on LLNL’s Blue Gene/L system and then discuss the impact of anticipated future large scale machine designs on Miranda’s performance. For the latter, we are currently seeing a trend towards multi- and many-core based parallel systems. In order to exploit such architectures efficiently, applications must make use of all cores within a CPU and consequently have to be able to deal with additional problems caused by this trend, including shared caches, increased memory contention and limited off-chip bandwidth. We investigate the impact of such architectures on Miranda and present preliminary results. We study the performance of Miranda on several current multi-core systems and show that, while not a problem just yet, future systems may pose substantial hurdles.