Periodic Reporting for period 2 - Inno4Scale (Innovative Algorithms for Applications on European Exascale Supercomputers)
Período documentado: 2024-07-01 hasta 2025-06-30
The EuroHPC Joint Undertaking (JU) is jointly funded by its members with a budget of around EUR 7 billion for the period 2021-2027. A large part of that funding is dedicated to the procurement of a European network of large- to extreme-scale supercomputers, with the inclusion of two exascale systems operational in the near future. In addition, the JU has funded research in application software tackling societal, scientific, economic or environmental challenges. However, exploiting ever faster computer systems poses critical challenges for many of the existing application software suites with a number of issues which include: the widening gap between sheer compute power and the lagging ability to move data sufficiently fast; finding sufficient parallelism in existing algorithms; mapping the computations onto very heterogeneous hardware, such as multi-core CPUs combined with graphics processing units (GPUs).
The Inno4scale programme, started in July 2023, managed the complete life-cycle of a funding programme to identify promising algorithmic developments for exascale computing based on their scientific and technological excellence and expected impact when used in important applications software. This has been instrumented through a competitive call based on the "Financial Support to Third Parties" mechanism.
After the necessary preparations (web, tools, etc.), the programme started with a thorough dissemination and promotion campaign of the call through the EuroCC and PRACE networks, as well as LinkedIn directed advertisements. With a total funding of 4,1M€, the call received 51 applications, of which two were rejected in the administrative evaluation. The remaining 49 underwent scientific evaluation, and 27 of them were ranked above the set scientific threshold (10/15). The funding was distributed to the proposals received following the evaluation ranking upon exhaustion of the available budget, applying only a 20% reduction of funding to the last one. Funding agreements were prepared with all innovations studies, and all of them started between February and March 2024, for a duration lasting until the end of February 2025. On a quarterly basis, all studies provided financial and technical reports for progress monitoring, and specific technical assistance was provided to those needing it. Two dissemination and promotion events wereorganised towards the end of the project, and the results of the innovation studies are presented on the Inno4scale website and a set of highlight success stories have been collected in a booklet which presents the range of potential impacts for scientific and industrial applications in a manner accessible for all stakeholders and the broader HPC ecosystem. The brochure will be distributed at major user-oriented fairs worldwide.
After the necessary preparations (web, tools, etc.), the programme started with a thorough dissemination and promotion campaign of the call through the EuroCC and PRACE networks, as well as LinkedIn directed advertisements. With a total funding of 4,1M€, the call received 51 applications, of which two were rejected in the administrative evaluation. The remaining 49 underwent scientific evaluation, and 27 of them were ranked above the set scientific threshold (10/15). The funding was distributed to the proposals received following the evaluation ranking upon exhaustion of the available budget, applying only a 20% reduction of funding to the last one. Funding agreements were prepared with all innovations studies, and all of them started between February and March 2024, for a duration lasting until the end of February 2025. On a quarterly basis, all studies provided financial and technical reports for progress monitoring, and specific technical assistance was provided to those needing it. Two dissemination and promotion events wereorganised towards the end of the project, and the results of the innovation studies are presented on the Inno4scale website and a set of highlight success stories have been collected in a booklet which presents the range of potential impacts for scientific and industrial applications in a manner accessible for all stakeholders and the broader HPC ecosystem. The brochure will be distributed at major user-oriented fairs worldwide.
Inno4scale’s 22 Innovation Studies worked on their new algorithmic ideas developing proof-of-concept implementations to expose the merits of their new approaches. The topics range from using machine learning in novel ways to speed up repetitive computations, via exploiting special structures in linear systems, adding the time dimension as a target for parallelism, reducing accuracy where it is not needed to novel ways to organize and distribute computational tasks to fully exploit all available hardware.
A number of studies tackled numerical Linear Algebra– both sparse and dense linear systems – which is a central component of an enormous range of HPC applications. Exemplary results were: reduced computations and communication, in particular synchronization, to speed up algebraic multigrid by up to 8x; a memory-efficient data structure for binary sparse matrices with over 11x reduction in memory size and up to 5x reduction in multiplication.
Computational Mechanics, and Computational Fluid Dynamics in particular, could be said to be a “classic” application field for HPC and was covered by multiple innovation studies but with a range of specific applications and high-impact innovations. Some examples of the results are the following: Mixed-precision approaches on GPU systems to enable new possibilities for high-fidelity simulation of hydrogen combustion targeting clean energy production. Asynchronous multi-GPU algorithms with unprecedented speed and energy efficiency to expand our understanding of cloud formation and development. Enhanced performance and robustness of general purpose CFD by including machine-learning methods into the core simulation methodology. Novel task-parallel multigrid solvers combined with optimized GPU executions for extreme-scale simulation of turbulence.
As the scale and complexity of applications increases, a key approach with a growing importance is uncertainty quantification (UQ). In many real-world situations, crucial data is noisy, incomplete and / or of limited accuracy – or simply unknown and needs to be estimated by use of simulation. One of the Inno4scale studies addressed such a situation in earthquake simulation in a principled and scalable way. Another developed a combination of a multilevel Monte-Carlo algorithm with a parallel-in-time approach and achieved over 90% of cost savings for UQ in electromagnetic real-time control applications.
Further application impacts from the Innovation Studies include: Progress towards overcoming obstacles in the use of computer applications to replace some aspects of experimental particle physics experiments. Addressing the feasibility limit of the typical preprocessing-simulation-postprocessing workflow at exascale by increasing the overall system use via the extraction of data co-processing tasks from a running simulation on a GPU and mapping them onto an otherwise idling CPU.
A number of studies tackled numerical Linear Algebra– both sparse and dense linear systems – which is a central component of an enormous range of HPC applications. Exemplary results were: reduced computations and communication, in particular synchronization, to speed up algebraic multigrid by up to 8x; a memory-efficient data structure for binary sparse matrices with over 11x reduction in memory size and up to 5x reduction in multiplication.
Computational Mechanics, and Computational Fluid Dynamics in particular, could be said to be a “classic” application field for HPC and was covered by multiple innovation studies but with a range of specific applications and high-impact innovations. Some examples of the results are the following: Mixed-precision approaches on GPU systems to enable new possibilities for high-fidelity simulation of hydrogen combustion targeting clean energy production. Asynchronous multi-GPU algorithms with unprecedented speed and energy efficiency to expand our understanding of cloud formation and development. Enhanced performance and robustness of general purpose CFD by including machine-learning methods into the core simulation methodology. Novel task-parallel multigrid solvers combined with optimized GPU executions for extreme-scale simulation of turbulence.
As the scale and complexity of applications increases, a key approach with a growing importance is uncertainty quantification (UQ). In many real-world situations, crucial data is noisy, incomplete and / or of limited accuracy – or simply unknown and needs to be estimated by use of simulation. One of the Inno4scale studies addressed such a situation in earthquake simulation in a principled and scalable way. Another developed a combination of a multilevel Monte-Carlo algorithm with a parallel-in-time approach and achieved over 90% of cost savings for UQ in electromagnetic real-time control applications.
Further application impacts from the Innovation Studies include: Progress towards overcoming obstacles in the use of computer applications to replace some aspects of experimental particle physics experiments. Addressing the feasibility limit of the typical preprocessing-simulation-postprocessing workflow at exascale by increasing the overall system use via the extraction of data co-processing tasks from a running simulation on a GPU and mapping them onto an otherwise idling CPU.