Challenges of DEM: I. Competing bottlenecks in parallelization of gas–solid flows

Abstract An analysis is performed on the computational efficiency of performing gas–solid simulations based on the open source code MFIX, which was recently parallelized by domain decomposition method [1]. Although it is well known that the DEM portion for solving solid-phase flow dominates the computational load, the computational time is impacted both by the total number of particles in a simulation and the particle size, since the former determines the number of force calculations needed in one time step while the latter directly affects the size of the time step. One natural constraint in all simulations is the amount of time the user is willing to wait (wall-clock time) for a given simulation. Another constraint unique to the parallel run is the maximum number of processors available for one simulation, the upper bound of which is limited by the ratio of system dimension (width) to particle size. To evaluate the relative importance of these two bottlenecks, we simulate a number of pseudo-2D fluidized beds filled with particles of increasing size, and in which the static bed height is kept equal to the bed width. For a given particle size, the maximum feasible system dimension can be determined separately via both constraints mentioned above, thereby revealing the limiting computational bottleneck for a given particle size. It is shown that with increasing particle size, the limiting bottleneck switches from the amount of time we are willing to wait until the completion of a simulation to the maximum number of cores (processors) allowable in parallel simulations. The critical particle size at which the transition occurs is dependent on the values of two constraints. The analysis is thus helpful in selecting optimum system sizes for simulations based on tolerable waiting times and available parallel computing capabilities.