Malleable model coupling with prediction

Abstract

Achieving ultra scalability in coupled multiphysics and multiscale models requires dynamic load balancing both within and between their constituent subsystems. Interconstituent dynamic load balance requires runtime resizing - or malleability - of subsystem processing element (PE) cohorts. We enhance the Malleable Model Coupling Toolkit's Load Balance Manager (LBM) to incorporate prediction of a coupled system's constituent computation times and coupled model global iteration time. The prediction system employs piecewise linear and cubic spline interpolation of timing measurements to guide constituent cohort resizing. Performance studies of the new LBM using a simplified coupled model test bed similar to a coupled climate model show dramatic improvement (77%) in the LBM's convergence rate.