Addressing future simulation needs will not only require new algorithms and simulation methodologies but also codes which can effectively use all the available super computing resources. This is challenging problem considering that even today the top 10 super computers in the world are divided between Sunway MPP, Intel Xeon Phi, NVIDIA GPUs, IBM BG/Q, and Intel Xeon processors. Considering this diverse set of hardware architecture designs, it is clear that programming models which provide performance portability are required to keep the software maintainable into the future. For C++ codes, Kokkos is such a programming model. Originated at Sandia National Laboratories, where it is the default threading model, it is establishing itself now in the wider HPC community. A large number of applications are actively working to convert to Kokkos at this point, including applications addressing classical molecular dynamics, low and high Mach fluid dynamics, structural mechanics, combustion and climate simulations. In this talk a short description of the Kokkos eco-system will be given in the context of upcoming hardware challenges, before presenting evidence from applications that the goal of performance portability can be achieved. The data will include performance evaluations of Molecular Dynamics simulations using the Reax and the SNAP  potential, which utilize increased compute capabilities on modern super computers to greatly enhance the accuracy of simulations.

Solving the performance portability issue with Kokkos

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.