Periodic Reporting for period 2 - ExaFLOW (Enabling Exascale Fluid Dynamics Simulations)
Période du rapport: 2017-04-01 au 2018-09-30
The main goal of the ExaFLOW project was to address current algorithmic bottlenecks to enable the use of accurate CFD codes for problems of practical engineering interest. The focus was on different simulation aspects including:
• Accurate error control and adaptive mesh refinement,
• Solver efficiency via scalable numerical methods and preconditioners,
• Strategies to ensure fault tolerance and resilience,
• Heterogeneous modelling to allow for different solution algorithms in different domains,
• Parallel input/output for extreme data, employing novel data reduction algorithms,
• Energy awareness of high-order methods.
All ExaFLOW innovations are clearly targeted to enhance the efficiency and exploitability of CFD applications on large-scale (exascale) systems. This directly relates to reduced time for simulation, the possibility to simulate larger problems, and thus improved competitiveness of European industry in areas such as Energy, Aerospace, Automotive, etc.
Indirectly, via the application areas, societal challenges such as Energy consumption, exploitation of novel forms of Energy, Climate, and even Health and Well-Being are being impacted through improved CFD models and simulations.
The techniques developed by the project have been implemented in representative CFD codes that show an excellent potential for reaching exascale performance. These selected codes capture two different types of high-order discretisation methods, namely the spectral/hp element discretisation (in Nek5000 and Nektar++, both for incompressible flow) and finite different methods (in NS3D and SBLI, both for compressible flow). Our engineering for exascale efforts have also take energy consumption into consideration. Finally, the results have been validated in a number of pilot applications of concrete practical importance, including a set of industrial (automotive use cases from McLaren Racing and ASCS ) and academic pilot applications.
• Accurate error control and adaptive mesh refinement,
• Solver efficiency via scalable numerical methods and preconditioners,
• Strategies to ensure fault tolerance and resilience,
• Heterogeneous modelling to allow for different solution algorithms in different domains,
• Parallel input/output for extreme data, employing novel data reduction algorithms,
• Energy awareness of high-order methods.
All ExaFLOW innovations are clearly targeted to enhance the efficiency and exploitability of CFD applications on large-scale (exascale) systems. This directly relates to reduced time for simulation, the possibility to simulate larger problems, and thus improved competitiveness of European industry in areas such as Energy, Aerospace, Automotive, etc.
Indirectly, via the application areas, societal challenges such as Energy consumption, exploitation of novel forms of Energy, Climate, and even Health and Well-Being are being impacted through improved CFD models and simulations.
The techniques developed by the project have been implemented in representative CFD codes that show an excellent potential for reaching exascale performance. These selected codes capture two different types of high-order discretisation methods, namely the spectral/hp element discretisation (in Nek5000 and Nektar++, both for incompressible flow) and finite different methods (in NS3D and SBLI, both for compressible flow). Our engineering for exascale efforts have also take energy consumption into consideration. Finally, the results have been validated in a number of pilot applications of concrete practical importance, including a set of industrial (automotive use cases from McLaren Racing and ASCS ) and academic pilot applications.
The ExaFLOW project addressed a number of objectives to facilitate extreme scale CFD:
We addressed the challenge of developing scalable adaptive methods, where error estimators drive an adaption process in order to make highly efficient use of large-scale computational resources without a priori knowledge of the flow solution. To improve solver efficiency, we have investigated new numerical methods (CG-HDG, overlapping Schwarz and Schwarz-multigrid preconditioners) and efficient ways of their implementation.
New code generation methods have been developed for efficient evaluation of stencil operation, and low latency communication kernels have been formulated to improve strong scalability limits. We have investigated how to improve the scalability of state-of-the-art spectral element methods with the development of communication reducing mixed discretisation methods and the formulation of various preconditioners suitable for adaptive methods. Furthermore, we advanced the field of error control for heterogenous modelling, allowing different regions of the flow to be accurately modelled with different approaches, reducing computational cost and increasing scalability. We have increased energy awareness of high order methods and developed methods to reduce energy with minimal impact on run-time.
Finally, we addressed the extreme data challenge by developing methods to reduce the volume of data that must be transferred from memory to disks by using filters for structure extraction and data reduction, i.e. transforming the large “raw” data to feature- or structure-based data which are orders of magnitude more compact. Together with new, efficient I/O schemes, development of fault tolerance algorithms to ensure resilience to hardware faults has significantly increased the exascale readiness of our pilot codes.
The results of the project have been widely disseminated in the Exascale and CFD community through a number of high profile publications, the organization of workshops and minisymposia at important community events, the project website that featured a regular blog on the project results, as well as white papers and press releases.
The results have been made available publicly via industry-friendly licenses and are being exploited in the project’s pilot codes, which impact a large user community directly, as well as by the project partners in their future developments.
We addressed the challenge of developing scalable adaptive methods, where error estimators drive an adaption process in order to make highly efficient use of large-scale computational resources without a priori knowledge of the flow solution. To improve solver efficiency, we have investigated new numerical methods (CG-HDG, overlapping Schwarz and Schwarz-multigrid preconditioners) and efficient ways of their implementation.
New code generation methods have been developed for efficient evaluation of stencil operation, and low latency communication kernels have been formulated to improve strong scalability limits. We have investigated how to improve the scalability of state-of-the-art spectral element methods with the development of communication reducing mixed discretisation methods and the formulation of various preconditioners suitable for adaptive methods. Furthermore, we advanced the field of error control for heterogenous modelling, allowing different regions of the flow to be accurately modelled with different approaches, reducing computational cost and increasing scalability. We have increased energy awareness of high order methods and developed methods to reduce energy with minimal impact on run-time.
Finally, we addressed the extreme data challenge by developing methods to reduce the volume of data that must be transferred from memory to disks by using filters for structure extraction and data reduction, i.e. transforming the large “raw” data to feature- or structure-based data which are orders of magnitude more compact. Together with new, efficient I/O schemes, development of fault tolerance algorithms to ensure resilience to hardware faults has significantly increased the exascale readiness of our pilot codes.
The results of the project have been widely disseminated in the Exascale and CFD community through a number of high profile publications, the organization of workshops and minisymposia at important community events, the project website that featured a regular blog on the project results, as well as white papers and press releases.
The results have been made available publicly via industry-friendly licenses and are being exploited in the project’s pilot codes, which impact a large user community directly, as well as by the project partners in their future developments.
The ExaFLOW project contributed to an extremely important area of HPC, both in academia and industry. CFD is in the heart of modern engineering and an indispensable tool for areas such as automotive, aerospace, energy, weather & climate, biotechnology, etc. A particularly important area here is turbulence, as about 10% of the energy use in the world is spent overcoming turbulent friction. Improvements in this area have thus a clear environmental and societal impact. CFD is also one of the areas that have a clear need, and potential, to reach exascale. The ExaFLOW project made important contributions towards enabling CFD simulations on exascale systems taking into account algorithmic concepts as well as highly-parallel and energy-efficient implementations.
Specifically, ExaFLOW expanded the state of the art in the areas of
• Mesh Adaptivity, Heterogeneous Modelling, and Resilience:
We addressed the challenge of developing scalable adaptive methods, where error estimators drive an adaption process in order to make highly efficient use of large-scale computational resources without a priori knowledge of the flow solution. To improve solver efficiency, we have investigated new numerical methods (CG-HDG, overlapping Schwarz and Schwarz-multigrid preconditioners) and efficient ways of their implementation.
• Strong scaling at Exascale
New code generation methods have been developed for efficient evaluation of stencil operation, and low latency communication kernels have been formulated to improve strong scalability limits. We have investigated how to improve the scalability of state-of-the-art spectral element methods with the development of communication reducing mixed discretisation methods and the formulation of various preconditioners suitable for adaptive methods. Furthermore, we advanced the field of error control for heterogenous modelling, allowing different regions of the flow to be accurately modelled with different approaches, reducing computational cost and increasing scalability. We have increased energy awareness of high order methods and developed methods to reduce energy with minimal impact on run-time.
• Extreme Data Challenge
We addressed the extreme data challenge by developing methods to reduce the volume of data that must be transferred from memory to disks by using filters for structure extraction and data reduction, i.e. transforming the large “raw” data to feature- or structure-based data which are orders of magnitude more compact. Together with new, efficient I/O schemes, development of fault tolerance algorithms to ensure resilience to hardware faults has significantly increased the exascale readiness of our pilot codes.
• Energy Efficient Algorithms
ExaFLOW investigated both energy efficient and energy aware algorithms, with the focus being on the former. In particular we have investigated mixed-precision algorithms, compiler and CPU friendly algorithms, the use of different programming models and frequency scaling.
Specifically, ExaFLOW expanded the state of the art in the areas of
• Mesh Adaptivity, Heterogeneous Modelling, and Resilience:
We addressed the challenge of developing scalable adaptive methods, where error estimators drive an adaption process in order to make highly efficient use of large-scale computational resources without a priori knowledge of the flow solution. To improve solver efficiency, we have investigated new numerical methods (CG-HDG, overlapping Schwarz and Schwarz-multigrid preconditioners) and efficient ways of their implementation.
• Strong scaling at Exascale
New code generation methods have been developed for efficient evaluation of stencil operation, and low latency communication kernels have been formulated to improve strong scalability limits. We have investigated how to improve the scalability of state-of-the-art spectral element methods with the development of communication reducing mixed discretisation methods and the formulation of various preconditioners suitable for adaptive methods. Furthermore, we advanced the field of error control for heterogenous modelling, allowing different regions of the flow to be accurately modelled with different approaches, reducing computational cost and increasing scalability. We have increased energy awareness of high order methods and developed methods to reduce energy with minimal impact on run-time.
• Extreme Data Challenge
We addressed the extreme data challenge by developing methods to reduce the volume of data that must be transferred from memory to disks by using filters for structure extraction and data reduction, i.e. transforming the large “raw” data to feature- or structure-based data which are orders of magnitude more compact. Together with new, efficient I/O schemes, development of fault tolerance algorithms to ensure resilience to hardware faults has significantly increased the exascale readiness of our pilot codes.
• Energy Efficient Algorithms
ExaFLOW investigated both energy efficient and energy aware algorithms, with the focus being on the former. In particular we have investigated mixed-precision algorithms, compiler and CPU friendly algorithms, the use of different programming models and frequency scaling.