The science of Geomorphology characterizes
processes that sculpt Earth's surface – including the associated transfers and
trajectories of mineral mass and chemical constituents, and resulting in the
formation and destruction of sedimentary deposits. LLEM
is a massively parallel computer model that simulates Earth Surface Dynamics
from mountains to continental margins.
Geomorphology began as a descriptive
science with qualitative conceptual
models for landscape evolution over larger temporal and spatial scales. In
recent decades the discipline has undergone a quantitative revolution whereby
many fundamental Earth surface processes are predictable utilizing calibrated
mathematical models. Synthesizing and upscaling such numerical constructs into
overarching ‘surface system’ perspectives is now accomplished quantitatively through Landscape
Evolution Models (LEMs). However, scales beyond relatively small landscapes or
simple formulae have proved computationally challenging for LEMs utilizing
single-threaded CPU-based algorithms with low memory bandwidth. LEMs often
avoid (or greatly simplify) lowland fluvial landscapes characterized by ‘morphodynamics’ that are fast-evolving and complex (eg., numerically intensive and massively interconnected)
compared to those for uplands rivers now represented with simplistic, steepest-descent,
single-node-wide, single-directional flow (unrealistic for large rivers that
migrate and bifurcate). Two centuries after von Humboldt, the father of environmental science,
emphasized the how Earth surface and biological systems were fundamentally characterized
by interconnectivity at all scales, it is unfortunate that most available environmental
models are written for computer architectures that have poor interconnectivity
when scaled up to larger simulations.
Recent, transformative advances in massively
parallel Graphics Processing Units (GPUs) can now facilitate holistic modeling
of processes, fluxes, and stratigraphy over extraordinary temporal and spatial
scales while efficiently representing a high level of system interconnectivity
– on the land, across continental shelves, and over millions of years. For
example, it is possible to model the Quaternary evolution of the central Amazon
basin, or 3.2 million km2 of the Sunda
Shelf (above image shows the Gulf of Thailand and the Mekong Delta at 488ka,
with Sea Level at -32 m).
LLEM (pronounced ‘Elm’) is a GPU-based Landscape-Linked Environmental Model that leverages massively parallel computer
architectures (of thousands to millions of GPU cores) to simulate a diverse
suite of geomorphic processes (for millions to billions of model nodes),
including:
1)
Mountain & uplands processes,
including soil formation, erosion, & (simple) landslides
2)
Fast multi-directional flux routing
with NIAGRA (Networking Incorporating
A Gpu-paRallel Algorithm)
3)
Water surface levelling & flows
that respond robustly to changing sea levels & other perturbations with HYDRA (HYDraulic paRallel Algoritm)
4)
Water & sediment partitioning (for
all topographic configurations) accounting for channel width and back-water
effects
5)
Particulate transport simulated
throughout the model domain, both fluvial & marine, and separately for
sand, silt, clay, & organic carbon fractions
6)
Detailed fluvial processes (meandering,
anastomosing, avulsions, bars, levees & ridges, back-swamps, &
floodplain channels)
7)
Stratigraphy fully accounted for all
nodes (stratigraphic ‘big data’ is load balanced between RAM and Optane drives)
8)
Vertical deformation fields, including
Glacial Isostatic Adjustment (GIA) & tectonic uplift
9)
Continental shelf processes (diffusive,
advective, & tidal), including waves for shelves
and lakes
10)
Mixture modeling for conservative
constituents & tracers [in testing]
11)
Shoreline and estuarine processes,
including prediction of palaeo-waves [in development]
12)
Reaction-Transport modeling for labile
biogeochemical constituents [in
development]
LLEM incorporates established mathematical
constructs from the published literature, synthesized within the framework of a
novel suite of massively parallel GPU-based algorithms. LLEM is designed to
simulate ‘Grand Scientific Challenges’ of morphodynamic
connectivity through ‘Source-to-Sink’ systems, and therefore employs an
adaptable and scalable architecture (with optimization of memory bandwidth,
cache size, and latency).
Geometric structure is a triangular irregular
network (TIN), allowing adaptable spatial configurations and density variations
for model nodes. While any Delaunay triangulation is possible, the fastest models
are D8 or D6 configurations – regular geometries that offer significant
increases in speed and other computational benefits. Flexible TIN structures
require look-up indexes (for neighbors) that represent a significant
calculations and memory bandwidth bottleneck for CPUs, single-threaded architectural
challenges that can be parallelized across thousands of GPU cores with
extraordinarily fast memory. While LLEM is exceptionally fast for TINs compared
to CPU models, it is even faster for regular meshes – current LLEM development
is focused on maximizing the advantages of hexagonal meshes.
Modern GPUs offer ~15 Tflops
(FP32) of computational power per chip
(each containing >5000 processors) and memory bandwidths of ~1 TB/s (~20x
faster than CPUs and ~200x faster than 5GB Infiniband
interconnects between traditional supercomputer blades). With 2-8 GPUs
installable per ‘blade’, ~120 Tflops of effective computation (eg.,
high bandwidth & low latency to shared memory necessary for simulations of
interconnected environmental systems) is possible for a specially built
workstation (figures that are increasing at a dramatic pace). A decade ago this
desk-top workstation would be the fastest supercomputer on Earth! LLEM also runs
well on custom ‘gamer’ systems (both desktop and portable) that feature
superior cooling and over-clocked GPUs.
Written from scratch in CUDA, C++, and Matlab with massively parallel GPU-optimized code
throughout, LLEM has the upscale potential to run across billions of cores on
GPU-based supercomputers such as TITAN (17 petaFLOPS)
or the upcoming SUMMIT (>200 petaFLOPS). LLEM is also positioned to take advantage of recent
advances in FP16 and tensor computations on Volta-class GPUs, which now offer
up to ~1 PetaFLOP of tensor performance for a single
desktop GPU-based system. Coding and model testing continues, with tens of
thousands of lines of code under active development.
This site is under development. Information and
videos will be posted as simulation of different landscapes continues …
although coding remains the priority.