ESCAPE, a major project in ECMWF’s Scalability Programme, held its kick-off meeting at the Centre on 1 and 2 October.
ECMWF co-ordinates this project, which is funded by the European Commission's Horizon 2020 funding framework under Future and Emerging Technologies – High-Performance Computing (FET–HPC). The meeting brought together 35 participants from the 12 partner organisations.
ESCAPE (Energy-efficient Scalable Algorithms for Weather Prediction at Exascale) supports the development of an accurate and scalable numerical forecast model at ECMWF. It will co-ordinate work on new and computationally efficient numerical model components and their adaptation to novel hardware architectures for use in global and regional weather prediction.
ESCAPE partners include representatives from the ECMWF Member States who share these parts of the computer code with ECMWF: Météo-France, the Danish Meteorological Institute (DMI, for the HIRLAM consortium) and the Royal Meteorological Institute of Belgium (RMI, for the ALADIN consortium). Other partners are the German national weather service (DWD) and MeteoSwiss, who have recently invested significantly at a national level in developing new models; Loughborough University; two computing centres in Ireland (ICHEC) and Poland (PSNC); and three computer hardware companies (ATOS/Bull, NVIDIA and Optalysys).
The kick-off meeting reviewed the proposal and laid out the work plan for the initial period of the three-year project. Since ESCAPE requires substantial collaboration on software development and diverse processor types, ECMWF will implement a common platform for exchanging and testing software.
Code adaptation challenge
The project is looking at novel solutions that ensure code portability to a variety of accelerator and processor types, while specifically testing these new concepts on Xeon-Phi, GPUs (graphics processing units), and optical processors.
One of the project partners, MeteoSwiss, has very recent experience of adapting its codes to hardware which combines more computing power with greater energy efficiency.
The new supercomputer used by the Swiss national meteorological service since September has a very high eight-to-two ratio of GPUs to CPUs (central processing units).
“We have developed and implemented this particular approach while other centres are still evaluating different options. We’ve shown that four to five years of development is a realistic time frame,” says Oliver Fuhrer, a senior scientist in the modelling group at MeteoSwiss.
Stan Posey, NVIDIA’s HPC Programme Manager for the Earth system modelling domain, agrees that accelerator devices such as GPUs offer great potential to make supercomputers more efficient.
“Accelerator devices are attached to CPUs and are typically used to accelerate numerical operations that lend themselves to a massively parallel architecture,” he explains.
Energy efficiency
NVIDIA, which leads the ESCAPE work package on hybrid computing, strives to keep the energy consumption of its GPUs constant as it increases the computing power.
Another project partner, Optalysys, aims to deliver supercomputer performance using an optical device powered from a standard mains supply.
CEO Nick New explains that some of the numerical operations used in weather forecasting, such as performing spectral transforms and matrix multiplication, can in principle be mapped to optical processes.
“This laser-based, low-power technology is based on well-established principles. It uses diffractive optics to perform operations such as Fourier transforms optically,” he says.
Different hardware options will be exploited through the partnership with the Center for Excellence in Parallel Programming hosted by ATOS/Bull, the third computer hardware partner in ESCAPE.
Peter Bauer, the head of ECMWF’s Scalability Programme, emphasizes the need for energy efficiency. “Software components will have to be profiled and evaluated in terms of computational cost and energy efficiency. Power usage to complete forecasts in terms of 'energy to solution' will be one of the key drivers for numerical weather prediction in the future,” he says.
Peter makes it clear that the results of ESCAPE will be widely shared. “ESCAPE links to the EU-funded Centre of Excellence for weather and climate prediction (ESiWACE). This will facilitate the sharing of the ESCAPE outcomes with the entire Earth system community."