Example Responsible Computing Plan

Overview

Questions

• What could a responsible computing plan for a research project using HPC look like

Objectives

• Provide a concrete example of an HPC responsible computing plan

Project overview

This project uses DFT calculations to understand how physical properties of a crystalline compound that is a candidate for novel solar panel technology change with temperature and pressure. Many DFT (parallel) calculations need run to sample the structure of the compound at different temperatures and pressures to explore phase space.

Responsible Computing Plan

Positive environmental impact statement

This project will help develop new solar panel materials that require less use of rare earth metals and use more sustainable materials reducing the environmental impact of mining activities (both in terms of resource use and in terms of clean water use). The resulting materials, if viable, should be cheaper to produce and have a longer lifetime accelerating the uptake of solar power. As described in the plan below, we will run the research project in a sustainable way demonstrating how such projects can be undertaken in a way that delivers the research goals in a way that minimises GHG emissions. We will make sure that discussion of this approach is highlighted in all research outputs from the project demonstrating leadership and championing this approach in our research community.

Project approach

Responsive plan

This plan will be reviewed at quarterly intervals throughout the project to ensure it is being followed by project members and to update the approach in light of project evolution.

Preparation and resource selection

We will use a quick, lower-resolution methods to sketch out phase space for the compound and then select the minimum number of points to run high resolution calculations to capture the regions of interest with the minimal resource use. Unfortunately, pre-existing datasets of calculations on this compound do not exist but we will use known experimental structures from crystallographic databases as starting structures where possible to reduce the number of iterations required to optimise structures.

Running the calculations responsibly

All calculation methods will be benchmarked on a small number of iterations to select the minimum number of cores/GPUs needed to run the calculations and give results in a reasonable timeframe. The high accuracy methods will be tested on a single point to ensure they produce valid results before the full campaign of calculations are launched – if problems are found, the method will be refined and re-benchmarked. We will prioritise minimising the job size over time to solution to fit within the wider research workflow so we are not computing quicker than results can be analysed. Initial planning suggests that as long as individual calculations complete within a 2 weeks of submission this should be fine to complete the project aims.

Measuring emissions from HPC use

The national HPC facility we plan to use publishes emissions data. We will use this to calculate the emissions associated with our use of the facility and convert it to a rate of bands simulated per kgCO2e. This metric will be evaluated for all testing on this system throughout the project and used to identify opportunities to become more emissions efficient. The emissions data will be published as part of all the research outputs to allow for better estimation of emissions (and potential improvements) for future projects using similar methods.

Data management

Each coarse-level calculation produces large (approx. 10 GB) wavefunction files that can be used to restart calculations if needed in addition to standard log output. From the initial coarse phase space survey these wavefunction files will be compressed and kept for the lifetime of the project to allow fine-grained calculations to be run efficiently from these starting points. They will be deleted once the research outputs from the project have been produced – they will be kept for this long to ensure reviewer comments can be addressed efficiently with additional calculations if needed. The fine-grained calculations produce larger wavefunction files (approx. 50 GB). We have produced a list of all properties that need to be calculated from each fine-grained calculation to try and ensure that calculations do not need to be re-run. All the wavefunction files from the fine-grained calculations will be compressed and stored in the same way as for the lower-resolution calculations and deleted once they are no longer possibly needed. All data will be stored in a single, shared location with a clear labelling scheme so all project members can access them as needed.

Project finalisation

We will make the datasets (including calculation input and output files) public via the Materials Cloud (materialscloud.org) using the ODC-by licence to enable their re-use for follow-on research. Once research output production has been finalised, we will delete the compressed wavefunction files that are no longer needed. Repositories containing analysis scripts and tools will be archived (e.g. to Zenodo) and made public with an MIT licence. All outputs resulting from the work will reference the public datasets and analysis repositories.

Key Points

• This demonstrates what a responsible computing plan could look like