Boosted computing increases European weather forecasting accuracy The biggest challenge for state-of-the-art numerical weather prediction arises from the need to simulate complex physical phenomena within tight production schedules. Software and hardware shortcomings are holding back weather and climate modelling. Climate Change and Environment © IgorZh, Shutterstock Current extreme-scale application software of weather and climate services isn’t very efficient on existing central processing unit (CPU)-type processors. It has about 5 % peak performance, mostly due to a lack of arithmetic intensity. The software also isn’t able to adapt to rapidly evolving options for new processor hardware mainly because of a lack of flexibility in mapping specific computational problems onto heterogeneous computing units. This problem is further exacerbated by other drivers for hardware development that aren’t necessarily ideal for weather and climate simulations. Boosting performance and energy-efficiency of weather and climate modelling The EU-funded ESCAPE project aimed to “restore this imbalance through actions that fundamentally reform Earth system modelling,” says coordinator Dr Peter Bauer. The project developed a holistic understanding of energy efficiency for extreme-scale applications using heterogeneous architectures, accelerators and special compute units. The project team developed and tested the concept of fundamental algorithmic building blocks called dwarfs. Dwarfs represent functional units in the forecasting model. They developed new algorithms specifically designed for better energy efficiency and improved portability. “Assessing numerical methods and algorithms for dwarfs rather than entire models reduces the complexity of the code,” explains Dr Bauer. “It enables high-performance computing (HPC) centres, research groups and hardware vendors to focus on specific aspects of performance for which code restructuring and adaptation to novel processor architectures is more straightforward.” Project partners then adapted and optimised the resulting dwarfs in terms of computing performance for different hardware architectures. For spectral transforms on CPUs, they achieved efficiency gains of up to 40 %. Code optimisation for graphics processing units (GPUs) delivered speed-up factors of about 10-50 on a single node, and again by a factor of 2-3 when deployed on multiple GPUs. The ESCAPE team also focused on domain-specific languages (DSLs). When adapted to GPUs with a DSL, a dwarf calculating the advection of air showed a speed improvement of a factor of two compared to the manually adapted version. They investigated a range of numerical methods exploiting multi-grid solvers and different types of spatial discretisation and time stepping. Improvements in predictive skills for weather and climate ESCAPE will impact European excellence for employing exascale HPC in helping to facilitate one of the largest societal impact areas: high-resolution weather forecasting. More precise forecasts in both time and space are critical for travel, health, work and safety. “The weather’s impact on society via forecasting and preparation has been reduced thanks to ESCAPE’s predictive skill advances,” Dr Bauer says. “By modifying numerical algorithms and using new programming models, substantial improvements to both weather and climate predictions will be possible and affordable, leading to reliable and timely forecasted warnings,” concludes Dr Bauer. “ESCAPE represents a huge step forward in weather and climate modelling.” The project directly benefits all European Centre for Medium-Range Weather Forecasts (ECMWF) members and cooperating countries. It will support both the Copernicus Atmosphere Monitoring Service and the Copernicus Climate Change Service that rely on ECMWF’s Integrated Forecasting System. Keywords ESCAPE, weather, climate, domain-specific languages (DSLs), dwarf, forecasting, modelling, high-performance computing (HPC), graphics processing units (GPUs), energy efficiency, exascale
