Change is afoot in high-performance computing (HPC) in meteorology, from increasingly heterogeneous architectures to alignment with machine-learning (ML) developments.
These issues and more will be at the centre of ECMWF's 20th Workshop on HPC in Meteorology from 9 to 13 October 2023 in the historic city of Bologna, Italy.
To reflect current trends, the theme of the event is ‘Diversifying HPC’. The workshop brings together experts in HPC from across national weather centres, academia, national supercomputing centres and industry.
Heterogeneous computing
One of the trends in HPC is a move away from large, monolithic computing systems. This includes the growing use of graphics processing units (GPUs) in addition to central processing units (CPUs).
“In the immediate future, we are looking at increasingly mixed technology, because we are reaching the limits of merely relying on an ever greater number of CPUs as you run into several physical constraints,” says Christine Kitchen, the Deputy Director of Computing at ECMWF.
There is also an explosion in cloud computing, with cloud computing providers now increasingly found in the mid-range of the Top500 list of the most powerful supercomputing systems.
“The way applications are traditionally built for clouds, using containers and hypervisors, is very different to how current HPC systems are run, so there’s increasing heterogeneity here, too,” says Christine.
“Finally, there is a spectrum of diversity in the communities we are supporting, from traditional researchers that use the applications in ECMWF’s Integrated Forecasting System (IFS) to the increasing use of machine learning in weather forecasting and adoption of the cloud.”
All these trends will be represented in keynote talks and presentations at the HPC in Meteorology workshop.
ECMWF’s current HPC facility in Bologna includes some GPU capabilities for research.
Hardware and software
Changes in hardware, such as from CPUs to GPUs, and from on-premises to cloud technologies will have to go hand in hand with changes in software. That is why the workshop will explore both hardware and software developments.
“Hardware is the fundamental infrastructure that we have to be able to support, but ultimately we need the software to effectively exploit new hardware,” says Ioan Hadade, team leader of HPC Applications.
“We’ll be looking at how our Member States are doing this in terms of restructuring existing code or introducing new components.”
The workshop will feature talks on both hardware and software developments.
Machine learning
Recently, articles announcing a number of machine-learning solutions trained to produce weather forecasts have made the headlines. This is an active area of research at ECMWF.
The systems work by using a dataset such as ECMWF’s ERA5 reanalysis to train global models for medium-range forecasting. Once trained, they can produce forecasts using far less computing power than traditional forecasts.
“We are planning to introduce machine learning into ECMWF’s weather forecasts, too, to complement traditional numerical weather prediction techniques,” Christine says.
How and to what extent that will be done are still open questions, which the workshop will help to address.
“An exciting period of change is coming up, and we should embrace this opportunity of increasing diversity,” Christine says. “Combining expertise from the traditional HPC community with the cloud community and machine-learning experts will be a good opportunity to break down some of the barriers that exist.”
More information
More information on the 20th Workshop on HPC in Meteorology can be found on ECMWF’s website. The submission date for abstracts for the workshop is 23 July 2023. Registration is open until 25 August 2023.