Dynamic Load Balancing for Malleable Model Coupling

Dynamic load balancing both within and between constituent subsystems is required to achieve ultrascalability in coupled multiphysics and multiscale models. Interconstituent dynamic load balancing requires runtime resizing-or malleability-of subsystem processing element (PE) cohorts. In our previous work, we developed and introduced the Malleable Model Coupling Toolkit with a load balance manager implementing and providing a runtime load-balancing algorithm using PE reallocation across subsystems. In this paper, we extend that work by adding the ability to adapt to coupled models that have changing loads during execution. We evaluate the algorithm through a synthetic coupled-model benchmark that uses the LogP performance model as applied to parallel LU decomposition.