2020-07-23 All-Hands Presentation Meeting Notes
Presenter: Peter Thornton
Title:
Progress toward a 1km offline land simulation capability for North America
Abstract:
With E3SM focusing part of its Phase 2 coupled simulation effort on simulations using a regionally-refined grid over North America, there is value in having an offline land science focus for the same region. One benefit of that approach would be the delivery of a “gold standard” set of results describing the surface weather and climate, as well as the energy, water, carbon, and biogeochemistry processes over the continent. These results could serve as a point of reference for coupled simulations, providing evaluation metrics for improvements in climate prediction skill when switching from standard model resolution to the regionally-refined North American nested grid. We know from previous offline land simulations that the quality of driving data has a major impact on the fidelity of process simulation outputs. Observational data resources over North America make it possible to derive very high-quality forcing information, at finer resolution than the proposed regional refined mesh. This allows aggregation of offline results to the coupled simulation grid, facilitating the construction of uncertainty ranges for the evaluation of the regionally-refined coupled results.
Over the past year a team at ORNL has developed the capability to execute these high resolution simulations, including development of a new sub-daily 1km surface weather forcing dataset for North America, and modification of ELM to make efficient use of GPU resources on Summit to carry out the large calculation and manage the IO streams. In this presentation we summarize the main efforts and provide a few illustrative examples of how the tasks have been accomplished. We use the OpenACC framework to add new capability for data movement and task parallelism on the Nvidia GPU devices on Summit, and we use the native MPI parallelism in E3SM to operate over multiple Summit nodes and multiple GPU devices per node. We highlight the use of a sophisticated functional unit testing framework to expedite the analysis, testing, and parallelization of ELM modules and sub-modules for GPU implementation. We provide current performance statistics and indicate ongoing work to optimize performance. A major conclusion is that the deep copy capability of OpenACC allows us to use the nested subgrid hierarchical data structure of ELM on GPUs with few modifications.
Date
Time
- PT: 8:30 am
- ET: 11:30 am
Call Info
- web session: https://global.gotomeeting.com/join/570361173
- call number: (571) 317-3122 Access Code: 570-361-173, If busy, use alternate number: (773) 945-1029
Joining from a video-conferencing room or system? Dial: 67.217.95.2##570361173 , Cisco devices: 570361173@67.217.95.2
Attendees
- Peter Thornton
- Ruby Leung
- Michael J Prather
- Stephen Price
- Renata McCoy
- Shaocheng Xie
- Susannah Burrows
- Sarat Sreepathi
- Steve Klein
- Dali Wang
- Darin Comeau
- Khachik Sargsyan
- Zeli Tan
- Forrest M. Hoffman (Unlicensed)
- Jiwen Fan (Unlicensed)
- Xubin Zeng
- Philip Jones
- Chang Liao (Unlicensed)
- Hailong Wang
- Katherine Smith
- Mathew Maltrud
- Nicole Jeffery
- Xylar Asay-Davis
- Phil Rasch (pnl.gov)
- Karthik Balaguru
- Xiaojuan Yang
- Kang, Shinhoo (Unlicensed)
- Holly Davis (Unlicensed)
- Andy Salinger
- Yilin Fang
- Chris Golaz
- Hyun-Gyu Kang
- Mingxuan Wu
- David C. Bader
- Mark Taylor
- Chris Terai
- Peter Schwartz
- LeAnn Conlon
- Sam Silva (Unlicensed)
- Carolyn Begeman
- Xujing Jia Davis
- Emil Constantinescu (Unlicensed)
- Jon Wolfe
- Matt Hoffman
- Jayesh Krishna
- He Wang (Unlicensed)
- Michael Brunke
- Bryce Harrop
- Salil Mahajan
- Beth Drewniak
- Wuyin Lin
- Ryan Forsyth
- Peter Caldwell
- Youngsung Kim
Presentation