New Algorithm Tracks Billions of Particles on Everyday Computers

Scientists have developed a new computational that dramatically improves the simulation of systems involving vast numbers of particles, such as clouds, industrial mixers, and combustion engines. This advancement, detailed in a recent paper, enables researchers to track up to 256 billion particles on high-performance computing clusters and handle 1.4 billion particles on a single workstation GPU, a significant leap over previous limits. By optimizing how computers process data from fluid flows and particle movements, the approach makes high-fidelity simulations more accessible and efficient, potentially accelerating discoveries in climate science, engineering, and beyond.

The key finding is that an asynchronous algorithm, named SCALE-TRACK, allows simultaneous calculation of fluid dynamics and particle tracking on different parts of a computer system, such as CPUs and GPUs, without the delays that typically slow down such simulations. The researchers demonstrated that this maintains accuracy comparable to conventional techniques while scaling efficiently to exascale-ready systems. In tests, it achieved near-ideal performance when tracking particles across up to 256 GPUs, with errors reduced to as low as 0.04% using a constant extrapolator-corrector , as shown in Figure 3 of the paper.

Ology involves decoupling the Eulerian (fluid) and Lagrangian (particle) phases of simulations, using independent domain decomposition and cache-friendly data structures to minimize communication overhead. SCALE-TRACK employs an extrapolator-corrector technique to handle asynchronous coupling, where sources from particles are estimated and corrected to avoid errors. The software is written in Julia and coupled with external computational fluid dynamics solvers like OpenFOAM, allowing flexibility in applications. Validation included comparisons against analytical solutions and OpenFOAM's built-in tracking, with indicating robust performance across various test cases.

Analysis from the paper shows that SCALE-TRACK outperforms existing s in scalability and efficiency. In strong scaling tests with 8 billion particles, the Lagrangian part scaled almost ideally, while the Eulerian part began to deviate at 640 ranks due to communication overhead, as depicted in Figure 7. In semi-weak scaling runs, tracking 256 billion particles on 256 GPUs maintained near-ideal performance, highlighting the algorithm's capability for large-scale simulations. Additionally, on a local workstation, SCALE-TRACK processed ten times more parcels than OpenFOAM while achieving a 2.7-fold faster time-to-solution and 2.5-fold better energy-to-solution, as detailed in the cloud chamber test case.

Of this research are broad, enabling more detailed simulations of natural and industrial processes without requiring massive computational resources. For example, it could improve cloud modeling for climate predictions or optimize inhaler designs in pharmaceuticals. The open-source release of SCALE-TRACK allows researchers worldwide to adopt this , potentially leading to faster innovations in fields reliant on multiphase flow simulations. By making exascale computing more accessible, it addresses key limitations in previous Euler-Lagrange implementations, such as synchronization barriers and load imbalances.

Limitations noted in the paper include deviations in temperature values when compared to OpenFOAM, possibly due to differences in tracking s, and the Eulerian part's scaling s at high core counts. Future work could enhance SCALE-TRACK with collision models, dynamic load balancing, and support for unstructured grids. The algorithm's reliance on external solvers may also introduce compatibility issues, though its design aims for agnostic coupling. Overall, while marks a significant advancement, further refinements are needed to fully exploit its potential in diverse scientific applications.