Uniform Load Balancing in PhysBAM ’ s Fluid Simulation via Threading

For this project, we focussed on simulating free surface flows on a single multiprocessor machine. Dividing the simulation domain into partitions, with a process per partition (as in the existing PhysBAM’s MPI based implementation) can result in a skewed work load across processors. For example, one process may be in charge of regions that mostly contain air, resulting in little actual work being done, when that process is scheduled. A naive way to overcome this is to have more partitions and processes, so that work gets distributed more evenly between processors. However, this can result in large amount of communication on chip, as all ghost values need to be synchronized. Also, large number of partitions does not work well with stages like projection, which use techniques such as preconditionining to solve implicit equations. We used threading, instead of MPI, to overcome problems like load imbalance and different requirements for stages like advection and projection.