Parallel SnowModel (v1.0): a parallel implementation of a distributed snow-evolution modeling system (SnowModel)

SnowModel, a spatially distributed snow-evolution modeling system, was parallelized using Coarray Fortran for high-performance computing architectures to allow high-resolution (1 m to hundreds of meters) simulations over large regional- to continental-scale domains. In the parallel algorithm, the mo...

Full description

Saved in:
Bibliographic Details
Published in:Geoscientific Model Development Vol. 17; no. 10; pp. 4135 - 4154
Main Authors: Mower, Ross, Gutmann, Ethan D., Liston, Glen E., Lundquist, Jessica, Rasmussen, Soren
Format: Journal Article
Language:English
Published: Katlenburg-Lindau Copernicus GmbH 22.05.2024
Copernicus Publications
Subjects:
Accuracy
Algorithms
Allocations
Atmospheric models
Blowing snow
Cores
Cryosphere
Datasets
Evolution
Freshwater resources
High resolution
Microprocessors
Modelling
Parallel processing
Physics
Regions
Scaling
Simulation
Snow
Solar radiation
Stream flow
Topography
Wind models
ISSN:1991-9603, 1991-962X, 1991-959X, 1991-9603, 1991-962X
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:SnowModel, a spatially distributed snow-evolution modeling system, was parallelized using Coarray Fortran for high-performance computing architectures to allow high-resolution (1 m to hundreds of meters) simulations over large regional- to continental-scale domains. In the parallel algorithm, the model domain was split into smaller rectangular sub-domains that are distributed over multiple processor cores using one-dimensional decomposition. All the memory allocations from the original code were reduced to the size of the local sub-domains, allowing each core to perform fewer computations and requiring less memory for each process. Most of the subroutines in SnowModel were simple to parallelize; however, there were certain physical processes, including blowing snow redistribution and components within the solar radiation and wind models, that required non-trivial parallelization using halo-exchange patterns. To validate the parallel algorithm and assess parallel scaling characteristics, high-resolution (100 m grid) simulations were performed over several western United States domains and over the contiguous United States (CONUS) for a year. The CONUS scaling experiment had approximately 70 % parallel efficiency; runtime decreased by a factor of 1.9 running on 1800 cores relative to 648 cores (the minimum number of cores that could be used to run such a large domain because of memory and time limitations). CONUS 100 m simulations were performed for 21 years (2000–2021) using 46 238 and 28 260 grid cells in the x and y dimensions, respectively. Each year was simulated using 1800 cores and took approximately 5 h to run.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1991-9603
1991-962X
1991-959X
1991-9603
1991-962X
DOI:10.5194/gmd-17-4135-2024