Search

This Te Ara Paerangi submission is from a small group of Research Software Engineers and advocates thereof. The primary focus of the submission is around how improvements to the public science sector could help develop our research workforce, particularly those in Research Software Engineering roles. Despite being critical to Aotearoa New Zealand's research success, these roles are considered non-traditional in a research context leading to RSEs suffering from many of the wprkforce challenges seen across the sector - precarity of funding, a lack of diversity, and a lack of career recognition.

/ INTRODUCTIONGPUs (Graphics Processing Units) have been used for many years to successfully accelerate scientific applications. The amount of work required to take advantage of GPU acceleration can vary significantly depending on the code you are using, the problem you are trying to solve, etc. In some cases it may be as simple as recompiling your code to enable GPU support; other cases may require modifying the code, for example using an API such as OpenACC to offload loops to the GPU, or even porting code to CUDA to take full advantage of the parallelisation offered by the GPU.NeSI’s consultancy service (https://www.nesi.org.nz/services/consultancy) can help researchers take advantage of the GPUs on our HPC systems. Here we will present an overview of NeSI’s GPU capability, including our recent investment into the latest NVIDIA A100 cards, and then highlight some case studies where we have helped researchers take advantage of GPU acceleration, to give the audience an idea of the amount of work involved and potential speedups. Some examples of recent work we have done in this area include: enabling GPU support in the molecular dynamics application NAMD to accelerate protein modelling simulations; using OpenACC to accelerate a code for computing the log-determinant of a matrix, a fundamental kernel of some data science applications; linking against cuBLAS to accelerate a tropical circulation model; and writing CUDA code to accelerate N-body simulations of the solar system.ABOUT THE AUTHORChris Scott is a Research Software Engineer for NeSI.Wolfgang Hayek is a Research Software Engineer for NeSI and NIWA.Alex Pletzer is a Research Software for NeSI.Maxime Rio is a Data Science Engineer for NeSI and NIWA.Georgina Rae is NeSI’s Science Engagement Manager.

This presentation, originally delivered in a webinar held on 6 May 2021, is about graphics processing units (GPUs). Wolfgang Hayek, NeSI and NIWA Research Software Engineer, explains what GPUs are, what type of research they are good for, and what you can do to get started using them.For a recording of the webinar click here: https://www.youtube.com/watch?v=PijLW7bpkUM

Many research areas from bioinformatics and genomics, to materials science, fundamental physics, earthquake simulation and weather/climate prediction are increasingly dependent on the availability of powerful computing platforms and deep software stacks. Unfortunately, scientific software too often runs at sub-optimal performance, sometimes reaching only a few percent of the maximum peak performance of the supercomputer. Small changes in code implementation details, the choice of compilers and libraries and adjustments in runtime environment have been shown to sometimes have a significant impact on code performance. By walking through some examples, we show how researchers were able to leverage NeSI’s free consultancy service to squeeze more performance out of their application, sometimes reducing the execution time by several factors, a win-win solution which benefits science and saves core hours.ABOUT THE AUTHOR(S) Alex Pletzer is research software engineer for NeSI at NIWA. Originally a physicist, Alex drifted towards high performance during a career that spans research in plasma physics, working for a private company in Colorado and supporting users at university in Pennsylvania. Chris Scott is research software engineer for NeSI at University of Auckland. Currently lead of the computational science team, Chris has a background in molecular dynamics, Monte Carlo methods, finite element analysis, visualisation and parallel computing Wolfgang Hayek is research software engineer for NeSI at NIWA and scientific programming group lead at NIWA. Wolfgang has expertise in radiative transfer modelling, fluid dynamics, data analysis and high performance computing Georgina Rae is engagement manager and, until recently, was team lead for the computational science team. Georgina has experience in food and plant research and has worked in the world of intellectual property and commercialisation

Containerisation is a form of virtualisation that has become very popular in the world of IT services and cloud computing, offering straightforward portability and deployment of software and services without having to install a complex set of dependencies. It has recently become available on High Performance Computing (HPC) systems through the popular Singularity software package. Singularity is a containerisation tool that is particularly suitable for HPC and scientific applications, featuring immutable software to support reproducibility of scientific results, as well as integration with HPC file systems, MPI, and more.This talk will outline the basic concepts of containerisation and discuss a recent NeSI consultancy project where a web server and database were containerised to process data on the Mahuika HPC. The project is now easily portable and can scale out to many cores, enabling very significant speed-ups.ABOUT THE AUTHOR(S) Wolfgang Hayek is a research software engineer at NeSI and NIWA, and group manager of NIWA’s scientific programming group, with many years of experience in scientific computing and HPC. Blair Bethwaite is solutions manager at NeSI; he has strong expertise in HPC, cloud computing, cloud architectures, and scientific computing Ben Roberts is an application support specialist at NeSI and has many years of experience in scientific computing and HPC

Parallel computing has become a necessity for a wide range of modern scientific computing problems, including data-oriented computing at large scale to achieve reasonable processing times. Implementing parallel computing can be challenging and time-consuming - APIs such as the Message Passing Interface (MPI) are powerful but can be hard to learn and implement. Dask is a popular toolkit for the Python programming language that addresses this issue. While requiring very little programming effort, it offers a variety of parallelisation paradigms, including work sharing via parallel function evaluation, task graphs, and direct integration with packages such as NumPy, Pandas, and Scikit-Learn. Dask can be used interactively and as a batch processing tool. The Dask-MPI package adds MPI as a parallelisation backend, enabling scalability and high throughput on high-performance computing (HPC) systems. In this presentation, I will introduce Dask, discuss some of its parallelisation mechanisms, and demonstrate how to use the MPI backend for batch processing. [Note: This presentation should precede Maxime Rio’s demo of using Dask with SciKit-learn in Jupyter notebooks as it will cover off the basics of Dask.] ABOUT THE AUTHOR(S)Wolfgang Hayek is a research software engineer at NeSI and NIWA, and group manager of NIWA’s scientific programming group, with many years of experience in scientific computing and HPC. Maxime is a data scientist at NeSI and NIWA. He enjoys helping researchers to analyse their data, from visualisation to probabilistic modelling. Chris Scott is a Research Software Engineer at NeSI with a background in scientific computing and HPC.

TensorFlow is a popular Python library for developing machine learning models, with a wide range of applications. Machine learning, and in particular deep learning, can be computationally very demanding. TensorFlow is therefore typically used with GPUs or specialised hardware. However, almost every modern computer comes with multiple CPU cores with considerable computational power. Running TensorFlow on multicore CPUs can be an attractive option, e.g., where a workflow is dominated by IO and faster computational hardware has less impact on runtime, or simply where no GPUs are available.This talk will discuss which TensorFlow package to choose, and how to optimise performance on multicore CPUs. We will also compare runtimes of training and inference tasks of a deep learning model between different CPU and GPU configurations as an example of a real-world application.

Motivated by the work of Guillot (2010), we present a semi-analytical formalism for calculating the temperature-pressure profiles in hot Jovian atmospheres which includes the effects of clouds/hazes and collision-induced absorption. Using the dual-band approximation, we assume that stellar irradiation and thermal emission from the hot Jupiter occur at distinct wavelengths ("shortwave" versus "longwave"). For a purely absorbing cloud/haze, we demonstrate its dual effect of cooling and warming the upper and lower atmosphere, respectively, which modifies, in a non-trivial manner, the condition for whether a temperature inversion is present in the upper atmosphere. The warming effect becomes more pronounced as the cloud/haze deck resides at greater depths. If it sits below the shortwave photosphere, the warming effect becomes either more subdued or ceases altogether. If shortwave scattering is present, its dual effect is to warm and cool the upper and lower atmosphere, respectively, thus counteracting the effects of enhanced longwave absorption by the cloud/haze. We make a tentative comparison of a 4-parameter model to the temperature-pressure data points inferred from the observations of HD 189733b and estimate that its Bond albedo is approximately 10%. Besides their utility in developing physical intuition, our semi-analytical models are a guide for the parameter space exploration of hot Jovian atmospheres via three-dimensional simulations of atmospheric circulation., Comment: Accepted by MNRAS. 18 pages, 8 figures. No changes made from last version

Monthly Notices of the Royal Astronomical Society, 420 (1), ISSN:0035-8711, ISSN:1365-2966, ISSN:1365-8711