Porting of DSMC to multi-GPUs using OpenACC

The Direct Simulation Monte Carlo has become the method of choice for studying gas flows characterized by variable rarefaction and non-equilibrium effects, rising interest in industry for simulating flows in micro-, and nano-electromechanical systems. However, rarefied gas dynamics represents an open research challenge from the computer science perspective, due to its computational expense compared to continuum computational fluid dynamics methods. Fortunately, over the last decade, high-performance computing has seen an exponential growth of performance. Actually, with the breakthrough of General-Purpose GPU computing, heterogeneous systems have become widely used for scientific computing, especially in large-scale clusters and supercomputers. Nonetheless, developing efficient, maintainable and portable applications for hybrid systems is, in general, a non-trivial task. Among the possible approaches, directive-based programming models, such as OpenACC, are considered the most promising for porting scientific codes to hybrid CPU/GPU systems, both for their simplicity and portability. This work is an attempt to port a simplified version of the fm dsmc code developed at FLOW Matters Consultancy B.V., a start-up company supporting this project, on a multi-GPU distributed hybrid system, such as Marconi100 hosted at CINECA, using OpenACC. Finally, we perform a detailed performance analysis of our DSMC application on Volta (NVIDIA V100 GPU) architecture based computing platform as well as a comparison with previous results obtained with x64 86 (Intel Xeon CPU) and ppc64le (IBM Power9 CPU) architectures.