Oliveira M1, Sinha M1, Yang E1
1ACCESS-NRI, , Australia
Biography:
Micael leads the Software Transformation Team at ACCESS-NRI, bringing the specialist software engineering skills required for the ACCESS modelling framework to take advantage of advances in computing hardware, and to position ACCESS for future changes in technology. Micael enjoys designing and developing complex pieces of software and enabling new science through software engineering best-practices and use of state-of-the-art hardware architectures.
Micael has a PhD in Computational Physics, from the University of Coimbra, Portugal, 2009. He is interested in all aspects of software development and high-performance computing applied to fundamental scientific research. He has a lot of experience in developing massively parallel scientific codes, and is a strong advocate for well written, modular, reusable, easy to maintain software.
Abstract:
Earth system models are some of the most computationally demanding scientific applications, routinely requiring thousands of CPU cores and substantial HPC resources. Being able to quickly evaluate the parallel performance and efficiency of climate models and to optimise their execution is thus an essential component of their development cycle. Given the complexity of the models, this poses a considerable challenge, even when using state-of-the-art tools, such as profilers specifically targeting HPC applications.
The Software Transformation team at the Australia's Earth Simulator, ACCESS-NRI, has been developing open-source tools to address this challenge across the ACCESS model suite running on HPC systems. The main purpose of these tools is to provide a lightweight, model-agnostic API that allows users to write reusable workflows for a variety of use-cases, from simple code profiling to more complex tasks, such as performing parallel scaling studies or improving MPI load-balancing.
This presentation will provide an overview of the tools' design and main features, highlighting how easy they are to extend to new climate models or other types of software. We will present some practical applications, focusing on a curated collection of Jupyter Notebooks that generate and publish parallel scaling data for several models. These notebooks have been useful in guiding parallelisation choices and providing researchers with data to be used in applications for merit allocation schemes, like the National Computational Merit Allocation Scheme (NCMAS). Because it addresses challenges common across research domains, this approach is of interest for software developers, HPC specialists and users in general.
