high performance computing on graphics processing units: hgpu.org

In recent years, researchers have employed a wide array of multi-physics computational tools, of varying sophistication, to simulate brownout conditions [1-3]. Among these tools, compressible high-fidelity Reynolds-Averaged Navier Stokes (RANS) solvers [3] depend the least on empirical assumptions. However, the high computational expense involved in RANS simulations of viscous, rotary environments, makes it less attractive compared to more efficient Free-Vortex Methods (FVM) or Vorticity Transport Models (VTM) . For this reason, large scale RANS solvers are usually run in parallel on multiple processors by using domain decomposition methods. For the large grids employed in brownout simulation, this strategy requires the use of large clusters and runtimes of the order of several days. Since their introduction in 2006, programmable graphics processing units (GPUs) have become ubiquitous owing to their high floating-point operation speeds, large memory bandwidths and low power consumption compared to multi-core CPUs. In the context of CFD applications, this architecture is particularly suited for data-parallel, fine-grain computations like explicit time-stepping, inviscid flux calculations etc. However, certain operations like implicit reconstruction schemes and implicit time-stepping can not be performed using fine-grain algorithms. In these cases, a careful restructuring of the algorithms, to exploit multiple levels of granularity, is required to ensure that accompanying performance penalties are not prohibitive. Preliminary results presented in this abstract using a GPU-based, 2D, compressible RANS solver suggest that significant speedup is achievable even when the computations involve the inversion of implicit systems. The primary goal of this research endeavor is to extend this solver to 3D for the purpose of simulating rotorcraft brownout.