Technical Papers Archive

Toward Continental Forest Dynamics Modeling at Individual-Tree Resolution on Exascale Systems


Authors: Xi Zhang and Bin Wang (Oak Ridge National Laboratory (ORNL)), Herman Shugart (University of Virginia), and Forrest Hoffman and Robert Patton (Oak Ridge National Laboratory (ORNL))

Abstract: Modeling continental-to-global forest dynamics at individual-tree resolution is essential for Earth system prediction but has remained computationally impractical because of cross-scale interactions among billions of trees in heterogeneous atmospheric and soil environments. We present GGap, an agent-based forest model designed for exascale systems. GGap employs a Site–Gap–Tree hierarchy that co-locates each site's ensemble on a single GPU to maximize data locality. It leverages two-level parallelism: site partitions run concurrently across MPI ranks, while GPU threads update agents within each rank. To avoid CPU–GPU synchronization, a multi-priority kernel pipeline executes sequential ecological phases entirely on-device, using grid barriers to enforce their order. On Frontier, the largest weak-scaling configuration simulates 10.25 billion agents on 2,048 GPUs at 96.2% parallel efficiency. A 1,000-year contiguous United States simulation across 1,424 sites and 235 species reproduces regional patterns of species composition, forest structure, productivity, and biomass, demonstrating ecologically meaningful continental-scale forest simulation.


Presentation: file

Back to Technical Papers Archive Listing