Scalability and Performance of Data-Parallel Pressure-Based Multigrid Methods for Viscous Flows

Abstract

A full-approximation storage multigrid method for solving the steady-state 2-dincompressible Navier–Stokes equations on staggered grids has been implemented in Fortran on the CM-5,using the array aliasing feature in CM-Fortran to avoid declaring fine-grid-sized arrays on all levels while still allowing a variable number of grid levels. Thus, the storage cost scales with the number of unknowns,allowing us to consider significantly larger problems than would otherwise be possible. Timings over a range of problem sizes and numbers of processors, up to 4096 × 4096 on 512 nodes, show that the smoothing procedure, a pressure-correction technique, is scalable and that the restriction and prolongation steps are nearly so. The performance obtained for the multigrid method is 333 Mflops out of the theoretical peak 4 Gflops on a 32-node CM-5. In comparison, a single-grid computation obtained 420 Mflops. The decrease is due to the inefficiency of the smoothing iterations on the coarse grid levels. W cycles cost much more and are much less efficient than V cycles, due to the increased contribution from the coarse grids. The convergence rate characteristics of the pressure-correction multigrid method are investigated in a Re = 5000 lid-driven cavity flow and a Re = 300 symmetric backward-facing step flow, using either a defect-correction scheme or a second-order upwind scheme. A heuristic technique relating the convergence tolerances for the coarse grids to the truncation error of the discretization has been found effective and robust. With second-order upwinding on all grid levels, a 5-level 320× 80 step flow solution was obtained in 20 V cycles, which corresponds to a smoothing rate of 0.7, and required 25s on a 32-node CM-5. Overall, the convergence rates obtained in the present work are comparable to the most competitive findings reported in the literature.