This deliverable reports on the profiling and benchmarking activities related to the development and optimisation of the SPACE codes. Each code is benchmarked at two key stages, the initial version at project inception (M1) and the current version (M30), using representative use cases from WP1 at scaled-up configurations, with scaling tests up to 64 nodes, to demonstrate the optimisation progress on the overall application runtime and on the kernels identified in D2.1. The second part of each code section covers the profiling activities across CPU and GPU architectures, including the audits completed in collaboration with the POP3 Centre of Excellence and the detailed runtime, kernel-level and architectural analyses carried out by the project partners to identify bottlenecks and guide further optimisation.
This deliverable reports on the efforts in Task 3.1 and Task 3.3 towards developing the visualisation and workflow tools proposed in the project, and coincides with Milestone M7: the provisional but fully functional integration of post-processing and visualisation tools. Guided by the use cases and requirements collected in D3.1, the document describes the first release of three visualisation use cases and one workflow use case: in-situ parallel visualisation through Hecuba; interactive, volume-based cinematic visualisation; interactive, distributed visualisation for HPC and cloud environments; and topology-aware workflows for modular Exascale data analysis. For each use case the tool development, results and work plan for the next release are presented.
This deliverable reports on the efforts in Task 3.3 towards developing the Machine Learning (ML) tools proposed in the project. The development is guided by the three ML use cases and user requirements identified in D3.1, selected for their relevance to the Astrophysics and Cosmology community and because they represent challenges in simulation development and data analysis that are common to other scientific domains. The document presents the first release of results for the three use cases: representation learning for explorative knowledge discovery in simulations; physics-informed reconstruction using synthetic observations of cosmological simulations; and predicting the effect of radiation on gas cooling rates on-the-fly in hydrodynamical simulations.
This deliverable reports and evaluates the activities implemented to enhance the energy efficiency of all SPACE applications, quantified as the ratio of performance (FLOPs measured through platform-specific performance counters) to power consumption. Following a static frequency tuning approach, where a single CPU/GPU frequency configuration is set at the beginning of each run, execution time, energy consumption and FLOPs are monitored and assessed. Building on D2.2, the report evaluates three new hardware platforms: the NVIDIA A100 GPU installed in the IT4Innovations Karolina supercomputer, the Intel Sapphire Rapids processor with DDR and HBM memory, and NVIDIA's ARM-based Grace CPU. The comparison of performance per watt shows that the NVIDIA Grace CPU is the most efficient CPU for all codes.
This deliverable reports on the optimisation activities performed by month M24 for the seven codes of the SPACE project. Each code partner describes its ongoing optimisation strategy, designed on the basis of the work reported in D1.2, D2.1 and D2.2, focused either on a specific kernel or on the code as a whole and covering vectorisation improvement, CPU optimisation, GPU porting and optimisation, and scalability improvements. On CPUs, most of the work relates to inter-domain MPI communication and hybrid MPI/OpenMP parallelisation; for GPUs, a broad range of porting approaches such as CUDA, Kokkos and OpenACC is presented. The gains obtained in CPU, GPU and scalability performance are assessed for each code, and future optimisation actions are outlined. This deliverable serves as proof that the project reached Milestone 6 at M24.
This deliverable reports on all optimisation activities and their impact on code performance. Each code describes the actions taken in optimising different kernels, their current status and the results obtained, with improvements ranging from significant speed-ups of the CPU implementation to the availability of a first fully functional GPU version. The second aspect of the document is the continuous integration set up by each code, with test cases suited to its algorithms and scientific applications, to improve code integrity. The report also provides the status of porting each code to the EuroHPC JU systems, where all codes have been ported to the majority of systems and some tested on all of them, together with benchmarking results where available.
This document outlines the current achievements and ongoing efforts within the SPACE CoE. The primary focus is on optimising and porting kernels or modules of the codes involved in SPACE CoE to GPU architectures and adapting them for efficient use on EuroHPC JU clusters. These advancements are crucial for addressing the increasing computational demands of large-scale, long-duration astrophysical simulations. The document provides detailed updates for each code involved in the project, covering the current status of GPU porting, optimisations, CI/CD implementation, and performance on EuroHPC JU clusters. All the seven codes have an open-source/public repository, allowing the progress within the code to be tracked in the respective repositories. The combined efforts in these areas demonstrate the project’s commitment to advancing HPC capabilities and supporting the computational needs of astrophysical research.
This document provides high-level descriptions for each code in the SPACE CoE. Within this code description the focus is on the main algorithms, which are used for each code. These variety of algorithms range from Smoothed-Particle Hydrodynamics, Adaptive-Mesh-Refinements, different Runge-Kutta schemes to Particle in Cell algorithms. In addition, the modules and kernels identified by each code for optimization are described as well as how the mini-apps will be created for each code.
The main goal of this report is to provide SPACE CoE codes scalability on JU systems, performance assessments and identification of the regions of the codes that can be potentially extracted as mini-applications or kernels and optimized in the following project activities. To evaluate the scalability and efficiency of specific performance aspects in the SPACE CoE parallel codes, a performance model and analysis methodology developed within the POP and POP2 Centres of Excellence is used.
This document encloses the details of the collaboration with other Centres of Excellence (as pointed by Euro-HPC and coordinated by CASTIEL2) as well as a number of dissemination and training activities. In this framework, SPACE will collaborate with National Competence Centres for High Performance Computing (NCC) and hosting entities in order to foster the network and get support for event organization.
This deliverable is a report on requirements and use cases for the SPACE Machine Learning and visualization tools, and topology-aware workflows for modular High Performance Computing (HPC) applications. The use cases and requirements have been collected to better identify post-processing analysis for the simulation data products and to integrate run-time modules suitable for coupling the exascale applications with visualization tools (such as VisIVO, Blender and Paraview) and Machine Learning techniques including representation learning, generative AI and Convolutional Neural Networks.
Energy efficiency is defined as the ratio of performance (floating-point operations per second) per Watt consumed by the application. In this deliverable, the energy efficiency of SPACE CoE applications has been evaluated with the aim of improving it by changing the selected power knobs of the underlying hardware.
This document illustrates several scientific use cases in astrophysics that demonstrate the power of data-driven research. The selected scientific cases have two main characteristics.
On the one hand, they represent current cutting-edge problems in the respective research fields and allow us to examine how different types of data, including observational, simulated, and experimental data, are used to address pressing questions in astrophysics. On the other hand, they pose relevant computational challenges to the CoE codes and stress their capability to scale in several respects.