Can emerging general purpose graphics processing unit (GPGPU) technology be used to mitigate computational burdens in environmental models?

Aerosol particles remain one of the most uncertain contributors to climate change and air quality. Gas-to-aerosol partitioning is key to determining the chemical composition and amount of aerosol particles, thus environmental impacts (e.g. the amount is critical to predicting air quality). Owing to the complexity and diversity of atmospheric aerosol components, quantification of the properties that determine their highly uncertain climatic and human health impacts requires the development of novel technological applications. The many thousands of individual aerosol components ensure that explicit manual calculation of these properties is laborious and time-consuming; the emergence of explicit automatic mechanism generation techniques predicting up to many millions of individual components.

Due to heavy computational demands associated with this level of complexity, this presents two broad problems when trying to develop appropriate modeling frameworks to assess true environmental impacts:

1) It is impossible to include full complexity representations of aerosol processes within large-scale frameworks, such as regional climate models. As a result, reduced complexity representations are developed with the inevitable tradeoff between accuracy and improved performance.
2) To determine whether a parameterization of aerosol processes is suitable it is necessary to first perform sensitivity studies, comparing full representaions with parameterizations under a wide variety of conditions. This requires considerable computational power and time.

Traditionally, computer processors have been single core. Recently this has evolved to several cores on a processor, typically 16 in an HPC server node. The graphics industry has been creating graphics cards (GPUs) with thousands of cores in order for games to have realistic effects. Recently, General Purpose GPUs (GPGPUs), although now commonly called just GPUs, have become available as "accelerators" for compute-intensive work. GPGPUs are available at a fraction of the cost of more traditional high performance computing (HPC) facilities and generally affordable (and thus accessible to) research groups.

The advent of GPGPU computing is a new and exciting technological development. Some vendors and discipline areas have begun porting some codes to GPGPUs, yet the atmospheric chemistry field has little/zero work in this area. In this project we propose to quantify the performance of state-of-the-art models of gas-to-aerosol partitioning, as a first example, using the newly emerging GPGPU paradigm against the more traditional CPU implementations. This pump-priming activity is designed to act as a springboard for more generalized potential improvements in computational efficiency of chemistry schemes in environmental models.

The successful outcome of this proposal will mean not only faster process models but that these could potentially be incorporated in to regional air quality & meteorological models, bringing higher accuracy and cost effectiveness to their solutions whilst improving their time-to-solution. As the emergence of GPU technology is relatively new, it is important lessons learned during this project will be shared by the broader research community, quantifying how the challenges of extracting near peak GPU performance were met. To this end we will use online facilities and informatics tools to ensure wider benefits are realised.

The proposed work uses developments in newly emerging GPU technology to attempt a reduction in fundamental computational burdens that currently compromise state of the art models of aerosol behaviour. Given the relatively low cost of GPU technology compared to more traditional HPC resources, this proof of concept project could help pave the way for heavily reduced costs surrounding development and application of environmental prognostic models used by policy makers.

The primary non-academic end-users of the proposed programme output in the UK would be the Met Office via existing links with the UKCA Climate-Chemistry-Community-Aerosol model, a joint NCAS-Met Office programme funded by NCAS, GMR and DEFRA. Policy decisions with respect to quantification and mitigation of the climate impacts of aerosol require policy-related model simulations with at least a rudimentary but physically-based representation of organic aerosol. Such model descriptions are unavailable & our investigations into development of new, efficient computing technologies might enable the inclusion of state-of-the-art aerosol frameworks within such schemes. Other international non-academic agencies conducting IPCC simulations would be best placed to use the same reduced complexity SOA formalisms as supplied to the Met Office. PI Topping continually develops reduced complexity modules within ongoing NERC programmes so is perfectly placed to drive potential inclusion of any newly developed GPU libraries. In addition, the code library will be made available publicly via the University of Manchester informatics portal.

Lessons learned from optimizing code on the GPGPU architecture will be disseminated via the University of Manchester GPU Club, a forum for researchers to feed back to major GPU developers such as AMD and NVIDIA which is run by Dr Bane.

A frequently updated blog hosted on the university website will announce new updates regarding the progress of this project. This will include an introduction to the emerging area of GPU technology and environmental modeling. We will also document new publications where appropriate. The incorporation of modeling tools improved in this project within an informatics suite developed under a separate NERC grant (NE/H002588/1), for which Dr Topping is PI, can be tailored to provide a useful online teaching resource to engage end-users. A final report will be made available in layman's terms and posted on the website at each Manchester, together with a downloadable poster which will also be available from our website.

Measuring online traffic through our data portal will provide a very useful indication on the success of engaging the wider academic and non-academic end users. In addition we will consider the project a success if we present at 3 separate science festivals, attend 1 UK / overseas conferences, presenting work at each, and publish 1 research papers and 1 general interest article during the timeline of the project.