Dynamic load balancing in parallel particle methods

Contemporary scientific simulations require vast computing power to solve complex problems in natural and life sciences. High-performance computing (HPC) enables fast execution of scientific simulation codes by parallelizing the computational workload across processing elements (PEs). This is usually done by decomposing the computational domain into smaller subdomains and assigning each subdomain to a PE. The subdomains encapsulate computational elements such as computational meshes or particles. The workload of each PE is defined by the number of operations done using these computational elements. In many scientific simulations, the initial load balance is not preserved during the course of simulation. Outside factors (e.g., congestion in the cluster network) as well as simulation dynamics (e.g., movement of particles in the computational domain) may alter the loads of PEs and may cause load imbalance. In such situations, overloaded PEs require more time to finish their tasks while underloaded PEs remain idle and wait for the overloaded PEs. This protracts simulation runtime and results in redundant energy consumption and wasting resources, which we would like to avoid to the extend possible. To handle these load imbalance situations efficiently, dynamic load balancing (DLB) techniques have been developed. In DLB, one is interested in making fast corrections to the load imbalance. As opposed to static load balancing, DLB tries to avoid a brand new subdomain-to-PE assignment.