07-30, 13:00–13:30 (UTC), Green Track
The ExaSGD (Optimizing Stochastic Grid Dynamics at Exascale) application is part of the Department of Energy's Exascale project (ECP). We show our challenges of finding optimization algorithms for GPUs and present our prototyping framework that written end-to-end in Julia.
The ExaSGD (Optimizing Stochastic Grid Dynamics at Exascale) application is part of the Department of Energy's Exascale project (ECP). We use Julia and JuMP for large-scale optimization on the current generation of supercomputers. However, the recent switch from CPU to GPU in the upcoming generation of supercomputers poses great challenges for classic optimization algorithms. Sparse algebra and linear solvers are known to have poor performance on GPUs. Our project looks into novel optimization algorithms that are applied to the alternate current optimal power flow (ACOPF) problem. Current solvers supported by JuMP largely rely on sparse linear solvers. Our goal is to find algorithms that rely on denser algebra that scales over the large number of cores that a GPU provides. We present a complete software stack from modeling to the linear solver that is written completely in Julia and provide an end-to-end overview of the challenges met both at the modeling level as well as at the algebra level. We use automatic differentiation to generate the derivative code for the GPUs and show performance results for the power flow problem (PF) that look promising to be extended to OPF.
Michel Schanen is Computational Engineer at the Mathematics and Computer Science division at the Argonne National Laboratory. His background is automatic differentiation and large-scale numerical simulations. His group's research revolves around optimization algorithms for power systems.