Many–Core Sustainability by Pragma Directives

Abstract

Many-core hardware is well adopted in scientific computing for a number of applications in an academic setting. Uncertainty about upcoming architectures and large development times for this hardware result in a modest acceptance in industry for commercial use. An upcoming turn from language-based many-core programming towards directive-based frameworks, similar to OpenMP, is an attempt to tackle these issues. We present a case study for a many-core acceleration of a large-scale commercial CFD solver by means of such frameworks. We achieved a local acceleration of up to 45 for hot spots with recent hardware but the global speedup remains below 2. The main obstacle for an efficient instrumentation is the design and the complexity of the original software. Further, restrictions given by the hardware and the frameworks exist. Based on the results we sketch a long term plan for a further acceleration.

Performance of a Code Migration for the Simulation of Supersonic Ejector Flow to SMP, MIC, and GPU Using OpenMP, OpenMP+LEO, and OpenACC Directives

“…Specifically for Computational Fluid Dynamics (CFD), applications of OpenMP have proven their effectiveness better than MPI [6]. Meanwhile, OpenACC, with a very similar methodology than OpenMP, has recently appeared as a viable alternative for specialized low-level languages, such as CUDA C and OpenCL for fluid dynamics GPU based applications, with very good results [7,8]. The granularity in OpenACC is considered fine for its origin linked to CUDA.…”