Infrastructure and Techniques to Accelerate Operational Weather Codes on Next-generation Supercomputers

Temporary Supervisor

Associate Professor Peter Strazdins

Description

Climate and weather models currently consume vast amounts of supercomputer time, with the most dominant component being the atmosphere. In order to make accurate long-range forecasts, BoM requires high resolution global atmosphere and ocean models. Similarly, with the ACCESS project performing large-scale climate simulations, the amount of usage is exploding. However, these models are complex software systems, with large amounts of legacy code. The primary consideration is to correctly encode the science for meaningful simulations; the secondary is performance, particularly on large-scale parallel computers.  State-of-the-art supercomputers are becoming increasingly complex,  with nodes not only being made of highly multiple traditional processing cores, but with multiple manycore accelerators.

The aim of the research would be to investigate techniques to support the efficient porting of key codes used in weather forecasting to supercomputers with heterogeneous nodes, that is multicore main processors augmented with accelerators such as NVIDIA's General Purpose Graphics Processing Units (GP-GPUs).

 

Goals

There are two levels to this project (A) investigating individual applications, and manually improving performance, and (B) a systematic evaluation of a gropup of applications, wiht the aim of (developing tools that support semi-)automatic code parallelization. 

Level A is suitable for advanced year coursework to Honours projects; level B is suited to Honours to HDR projects.

The goals of this project include (1) analyzing and developing an understanding of the performance and scaling behavior of a selection of operational weather codes, (2) exploring and evaluating new opportunities and techniques for parallelization, in particular wiht the use of GPU accelerators. . At level B, this includes the use of new programming paradigms and code transformation tools. Work on supporting infrastructure includes automated methods to reverse-engineer test harnesses (correctness and performance) for selected performance-critical subroutines and generate kernels for them, methods to automatically refactor the codes for the desired target node architecture, and tools to reliably predict the performance of these codes on future accelerator designs. This infrastructure and techniques can be used to benefit many applications areas, but the project will aim to prove them on selected operational codes of interest to the Bureau of Meteorology and NVIDIA.

Requirements

(PhD Level) An Honours degree in computer or computational science or equivalent. Some background in high performance computing  is highly desirable. Experience in code transformations tools would be ideal.

(Coursework project and Honours level)  at least two years study in Computer Science, Some background in high performance computing  is highly desirable

This is a computer science project, knowledge in weather and earth sciences is not required.

Background Literature

Fan Yu, Peter E. Strazdins, Joerg Henrichs and Tim F. Pugh. Shared Memory and GPU Parallelization of an Operational Atmospheric Transportand Dispersion Application, 2019 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) pp 729-738, IEEE, May 2019.

Peter E. Strazdins, Margaret Kahn, Joerg Henrichs, Tim Pugh and Mike Rezny, Profiling Methodology and Performance Tuning of the Met Office Unified Model for Weather and Climate Simulations , Proceedings of the 2011 IEEE International Parallel & Distributed Processing Symposium Workshops, Anchorage, May 2011, pp1317-1326.

Gain

Weather science is of increasing importance, and the with it the need to perform efficient and meaningful simulations, especially for medium-term forecasts. This project represents an opportunity to join and make a significant contribution with an international team working in computer and earth sciences.

Over 2017-18, Honours student Fan Yu achieved the first OpenMP and CUDA parallelization of the widely used HYPSLIT particle tracking model (https://www.ready.noaa.gov/HYSPLIT.php), widely used around the world for tracking pollutant dispersion (e.g. from bushfires or the Fukujima 2011 disaster). Fan's accelerated  codes are now being used operationally in two international projects ( https://cecs.anu.edu.au/news/new-weather-code-sunn...), but there is scope for up to an honours project on further improvements.

 

Updated:  1 June 2019/Responsible Officer:  Dean, CECS/Page Contact:  CECS Marketing