Supercomputers are used for high calculation-intensive tasks. Typical applications relate to solve problems involving quantum physics, weather forecast, climate research, molecular modeling, or physical simulations
Biological macromolecules, polymers, and crystal require to compute chemical compound structures and properties. Designing green airplanes needs simulations in wind tunnels.
Advancing nuclear fusion research or understanding detonation of nuclear weapons force simulations.
A particular class of problems, known as Grand Challenge problems, are problems whose full solution requires semi-infinite computing resources.
Industrial and engineering applications
Many industries such as aircraft, oil & gas, transportation, use digital models to quickly evaluate and improve their products or processes. It is critical to save costs, improve quality, reach Time To Market to remain competitive.Virtual prototyping and large-scale data modelling resource are the typical industrial and engineering applications that require high performance computing.
For some domains, Exaflop capacities are needed to solve industrial problems:
- Aeronautics: Improved predictions of complex flow phenomena around full aircraft configurations with advanced physical modeling and increased resolution, multi-disciplinary analysis and design, real time simulation of maneuvering aircraft, aerodynamic and aero elastic data prediction, Exaflop not the ultimate goal, need at least Zetaflop for LES of complete aircraft. Many ―farming‖ applications (almost independent simulations)
- Seismic, Oil &Gas: Largely embarrassingly parallel, major problems are programming model, memory access and data management, need Zetaflop for full inverse problem
- Engineering: Optimization, Monte Carlo type simulations, LES/RANS/DNS … (main problem: equations themselves, physics, coupling, …)
For these domains, production problems will be solved by “farming” applications.
EESI1 studies have identified impacts of exascale in several sectors.
As examples, in the energy domain, the energy process needs to continuously improve its environmental impact (thermal discharge of nuclear/thermal power plants, long term geomorphology in rivers due to hydropower, water or air quality, …). This can be achieved by using multi-physics, multi-scale, complex three- dimensional time-dependent models
Improving the efficiency of the search for new oil reservoirs, including non-traditional ressources, can only be done through very advanced wave- propagation models and image processing methods.
In transportation, future systems will have to meet the constantly increasing needs of European citizens for travel and transport, as well as the strong requirements to preserve the environment and quality of life. By 2050, building green aircrafts and gain certification in short timeframe, will require to reduce key indicators:
- CO2 by 75%
- a NOx by 90%
- perceived aircraft noise by 65%
- accident rate by 80%
The main challenge facing nuclear industry is, today more than ever, to design safe nuclear power plants. HPC and in particular Exaflop computing will contribute to improve plant design, involving the use of multi-physics, multi-scale, complex three- dimensional time-dependent models.
Nevertheless access to HPC resources is a challenge. While some large companies have their own systems, eg Total, many companies do not have access to such resource, either because of the complexity, price, or lack of information.
Path: In order to raise awareness and foster innovation and competitiveness of SMEs through HPC, some national initiatives have been launched, such as SHAPE project of PRACE or Fortissimo EC funded project.
For aerospace, green objectives require to flight-test a virtual aircraft with all its multi-disciplinary interactions, with guaranteed accuracy. This means to couple LES simulations at all the different stages of turbines, to develop both simulation codes as well as flexible coupler and optimised coupling techniques.
Roadmap: investments into large petascale HPC systems have started in Europe by companies like Total or Airbus, becoming HPC bigger users in their domain.
After success in combustion, with a first jet noise simulation on one million core in collaboration with Aachen University and PRACE, aerospace companies set ambitious roadmaps for developing next generation of gas turbines by 2020.
Automotive has a strong use acceleration of advance HPC for crash analysis or steel replacement by composite.
The European industrial roadmaps have been challenged against US competitors and the results are strongly coherent.
Weather, Climatology and Solid Earth sciences (WCES)
Sciences related to Earth encompass a wide range of disciplines from the study of the atmosphere, the oceans, the biosphere to issues related to the solid part of the planet.
A key trend in this domain is the move from current climate models towards full Earth System Models.
Today it is clear that models must also include more sophisticated representations of non-physical processes and subsystems, which are of major importance for long-term climate development, like the carbon cycle. Scientists are strongly interested to know the sensitivity of predictions not only to unresolved physical processes (as, e.g., the cloud feedbacks mentioned above), but also to non- physical ones, like those related to biology and chemistry (including, for example, those involving the land surfaces, and -greenhouse gases reactions).
It should be noted that including the representation of biogeochemical cycles using different biochemical tracers and aerosols typically increase computing time by a factor of 5 to 20 (depending on the complexity of the parameterizations and the number of tracers). An increase of computing power by a factor of 5 to 20 is then required to better account for the complexity of the system.
Earth sciences are multiple, from weather prediction to air quality and to climate change forecast, from ocean prediction to natural hazards, such as seismic and volcanic hazards. Among all these issues WCES particularly requires exascale capacity in two domains:
- weather and climate, which share some similarities
- solid earth
Challenges: speaking of climate change, beyond greenhouse warming analysis, it is becoming crucial to shift from predictions to forecasts, ie changing scale of time from yearly to ten-yearly and learning evolution from initial ocean state to future.
Studying thermal convection, from the lower to the upper layers of the atmosphere, requires «subgrid scale parametrizations». Uncertainty of such modelisation will directly impact the quality of results of climate simulations.
Climate simulation will require an estimated grid size slightly smaller or equal to ca. 1 kilometer. This is a major challenge in terms of volume of data to be processed.
By-en-large, challenges impacting the two domains can be summarized as follows:
- Scientific data management: visualization, parallel files systems, parallelization strategies and frameworks for data analysis, programming models for big data, data compression, data mining, data quality, data provenance, metadata management, etc.
- Earth Sciences Modelling: finer resolution, development, evaluation of parameterization well adapted to those fine resolution, fast run for lower-resolution models, uncertainty quantification, quality control, new approaches (e.g., new grids and new dynamical cores), etc.
- Cross-cutting issues: multi-physics coupling software for 105-106 cores, resilience, and performance of individual modules and of complete coupled systems, parallel I/O
Path: improving scalability of applications, developing dynamical cores, coupling multiple codes, assessing uncertainties of the models.
Roadmap: Climate modelling requires very high spatial resolution for reaching satisfactory simulation of cloud and convection processes. As a step forward to reaching the longer-term, 1-km resolution,objective, the image shown in this section is a simulation at the spatial resolution of 25 km. It has been performed during the UPSCALE project using the Met Office Global Atmosphere Unified Model on the PrACE Tier0 machine Hermit
Fundamental sciences covers nuclear physics, laser- plasma physics, nuclear fusion, quantum chemistry, soft matter physics, materials science and astrophysics/cosmology. Those sciences strongly rely on fast algorithms and efficient software implementations while tracking developments in hardware technology.
There is no clear opinion yet, whether the transition from peta to exascale can be achieved smoothly from current implementations or whether a revolutionary step is necessary.
Challenges include programming models, communication schemes, memory management, algorithms including fault tolerance and energy efficiency.
Europe has a very strong position in the global scientific community and is leading in several fields of astrophysics/cosmology, nuclear/hadron physics, fusion research and materials sciences. Software development is very advanced in those domains.
Exaflop capacities in 2020 means being able to deliver a Petaflop in a box for $200 000 and 20 kW of power consumption. This represent a huge impact for those, academic, industrial, large and small structures, including SMEs, that will be able to take advantage of “exascale” technology, not just for few heroes applications
The needs pose crucial issues for handling very large amounts of data to be processed on the fly. In this area, exascale definitely means, Exaflop and Exabytes. The challenge is twofold, to compute at the speed of 1018 floating point operations per second AND the capacity to manage, explore, extract information and knowledge of data set size of 1018Bytes
Path: Developing multiscale software frameworks would position Europe ahead of competition.
Roadmap: massive simulations in the field of dark matter or collisions between galaxies have been performed by US, Japanese or European research teams.
GYSELA and PEPC fusion codes have been successfully scaled out to the full JUQUEEN machine at Juelich (half a billion cores using 1,835,008 threads)
Life sciences and health
The Life Science community is very diverse. Experimentalists/biologists strongly depends on computational results. Computational biologists or bio-informatics depend heavily on HPC resources to understand the mechanisms of living systems.
Research in the Life Sciences generates knowledge with a very clear and direct impact on our society. HPC is of great importance particularly in research related to health and biotechnology sector, that beyond big companies encompasses 2,100 biotechnology companies in Europe.
Challenges: Simulation will reduce costs, time to market and animal experimentation. In the medium to longer term, simulation will have a major impact on public health, providing insights into the cause of diseases and allowing the development of new diagnostic tools and treatments.
With the recent advances in this area (e.g. next generation of DNA sequencing instruments) the generated data is becoming larger and more complex. Appropriate computer resources need to process sequencing data at exascale, i.e handling ExaBytes of data.
In contrast to other communities there are no universal computer packages and software evolves very fast to adapt to new instruments.
The problems faced by scientists working in molecular simulations and genomics are also very different, as are the computer algorithms used.
The importance of having fast and flexible access to very large computational resources is crucial in the many fields of Life Sciences and the lack of suitable computers can block entire projects
Path: beginning of 2000, the Human Genome Project was an international flagship project that took several months of CPU time using a hundred-Gigaflop computer with one terabyte of secondary data storage. Today genomic sequencing is no longer a scientific milestone, but a powerful tool for the treatment of diseases, able to deliver results in days, while the patient is still alive
In coming years, sequencing instrument vendors expect to decrease costs by one to two orders of magnitude, with the objective of sequencing a human genome for $1000.
Roadmap: The European Commission has announced in early 2013 the launch of 2 flagship projects: Human Brain Project and Graphene (1B€ during 10 years per project) that will use PRACE and GIANT infrastructures.
PRACE has 6 petascale systems in France, Germany, Italy and Spain offering a cumulated performance of more than 15 PFlops. 12 companies, from SME to large size, have used the services.
The notion of “disruptive technology” implies the meaning of “breaking traditions”.
Challenges: in the context of application domains using exascale computing it is a challenge by itself to classify disruptiveness, since:
- exascale technologies are still under discussion
- “applications” refer to actual scenarios. Thus disruptive technologies must show a tangible potential to be successful.
Therefore, the selection of discussion topics is not exhaustive. It is based on fields where some development has already been achieved though not yet mature. They bare risks and a high potential at the same time, for a very large class of applications.
Discussion topics have an inherently interdisciplinary and multi-disciplinary feature. They often include a variety of length- and time-scales and a diversity of different methods, e.g. stochastics vs deterministics, lattice vs mesh-less methods, optimization, dynamics vs sampling methods including multiscale methodologies.
Very complex networks exist in nature and technological applications, which form a type of self-organized, synthesized or genetically determined structure. The function of such systems is difficult to recast in a straightforward and deterministic way.
In the end, various types of problems relates to the complexity of the problems itself.
Path: break-throughs can be reached if methodological and algorithmic advances are further progressed.
Multiscale techniques can be considered versatile method to characterise complex systems over various time- and length-scales. However, multiscale often implies multi-numerics, i.e. coupling different methods which have their own stability constraints and level of accuracy and a concurrent coupling of methods might not guarantee a possible error control.
Socio-economic or environmental problems, like climate change, neurological diseases, efficient energy-production and -usage or market crashes, can be understood using complex networks or functions within networks. This would require to address code couplers, training, Big Data, mini applications, ultra scalable solvers as well as uncertainty quantification & Optimization methods and tools.