Authors: Shan Jiang, Hikmet Alperen Aydin, and George Biros (University of Texas at Austin)
Abstract: Certification of mission-critical 3D-printed components can take over a year because it is sensitive to process parameters. Simulation could shorten this, but 3D printing spans vast spatial and temporal scales. We isolate one key physics challenge: predicting the temperature field. We introduce JIFFY, a parallel-in-time solver for heat transfer with a moving point source. Given the source path and an approximate Green's function, we construct a directed acyclic graph encoding space-time dependencies, decompose it into spatial path segments processed in parallel, and resolve remaining dependencies with a few iterations local in space and time. JIFFY achieves near-ideal weak scaling and good strong scaling. Our Python implementation (GPU+MPI) simulates a multilayer print on 64 GPUs in under 30 minutes at an effective uniform-grid resolution exceeding tens of billions of points and hundreds of thousands of time steps; a state-of-the-art single-GPU adaptively refined solver without temporal parallelism would need over a day.
Presentation: file