Exascale Computing Project (ECP)–funded researchers have demonstrated a massively parallel, scalable system for simulating physical behaviors of materials undergoing complex topological changes, collisions, and large deformations using multi-GPU supercomputers such as Summit at Oak Ridge National Laboratory. Their work supports ECP’s ExaAM project, which focuses on using leadership-class computing to model additive manufacturing processes such as 3D printing with metals. The research was published in the July 2020 ACM Transactions on Graphics.
Theirs is the first work aimed at leveraging hybrid computing architectures for Material Point Method (MPM) while using a sparse data structure concept to reduce memory footprint. MPM is being used in additive manufacturing simulations to model the complex interactions of metals and lasers in processes where material changes phase and moves dynamically.
The researchers’ method achieves over 100× per-time-step speedup, highlighting exciting possibilities for using MPM simulations in computer graphics and computational science; 2× acceleration on a single GPU, which achieves almost-linear scaling to multi-GPU systems; and near-perfect weak and strong scaling with 4 GPUs, enabling performant and large-scale simulations. They demonstrated their method’s effectiveness with extensive benchmarks, evaluations, and dynamic simulations with elastoplasticity, granular media, and fluid dynamics.
Based on the Material Point Method, their system introduces (1) a new, specialized Array-of-Structs-of-Array particle data structure that promotes coalesced memory patterns on the GPU and eliminates the need for complex atomic operations on the memory hierarchy when writing particle data to the grid; (2) a kernel fusion approach using a Grid-to-Particles-to-Grid scheme that reduces kernel launches, improves latency, and reduces the amount of global memory needed to store particle data; and (3) optimized algorithmic designs that allow for efficient sparse grids in a shared-memory context.
The researchers’ results show that MPM can be deployed efficiently on exascale-class computing architectures for problems where the physical material may be sparsely distributed. Their method is expected to map well to future DOE exascale computers including Frontier.
Wang, Xinlei, Yuxing Qiu, Stuart R. Slattery, Yu Fang, Minchen Li, Song-Chun Zhu, Yixin Zhu, Min Tang, Dinesh Manocha, and Chenfanfu Jiang. 2020. “A Massively Parallel and Scalable Multi-CPU Material Point Method.” ACM Transactions on Graphics 39 (4) (July 8). doi:10.1145/3386569.3392442. http://dx.doi.org/10.1145/3386569.3392442.