Successful Collaboration with CoEs
We cooperate very successfully with other EuroHPC JU Centres of Excellence in HPC. We do this by providing POP performance assessments and second-level services, and by disseminating results achieved through joint events or coordinated publications. These services can take the form of assessment campaigns, in which POP analyses all application codes belonging to a CoE within a specified timeframe, or single performance services for a specific CoE lighthouse code.
A highlight of the first 18 months of the project was the successful ChEESE measurement campaign, which was documented in a joint webinar. Another success story was the Analysis and Optimization of the Vlasiator code at Production Scale, used in the Plasma-PEPSC CoE. The Project section of our Resource for Co-design website provides a list of CoEs, along with the list of outcomes of services we performed for them. Two workshops were held at HiPEAC25, which were co-organised with ChEESE, EoCoE, ESiWACE3, EXCELLERAT, MaX, MultiXscale and SPACE, to disseminate CoE success stories and project outcomes. Finally, we organised a workshop on "Readiness of HPC Extreme-Scale Applications" at ISC 2024, ISC 2025 and ISC 2026 with contributions from CEEC, HiDALGO2, ChEESE, MaX, ESiWACE, dealii-X, EXCELLERAT, MultiXScale, SPACE and EoCoE-III.
Effective Training in Performance Analysis Methods and Tools
We continue to successfully organise and carry out training activities related to POP performance analysis metrics, methods, and tools. Project partners BSC, HLRS, INESC-ID, IT4I, FZJ, RWTH and UVSQ combine their training activities within the framework of the Virtual Institute – High Productivity Supercomputing (VI-HPS). Workshops and tutorials are organised in close collaboration with the host organisation to meet the specific needs of the audience. Training events can be tuning workshops lasting three to five days or tutorials conducted in collaboration with HPC-related conferences. Tuning Workshops are our main training initiative, in which around 15 to 30 participants receive instruction and guidance on applying POP tools to their own parallel application codes, along with advice on potential optimisations. Feedback from participants helps to direct the development of tools towards user needs, as well as improving tool documentation and ease of use. Coordinated training materials on tools are available, with an emphasis on hands-on exercises using POP tools individually and interoperably. The training activities are supplemneted by one-hour webinars on HPC performance tools and parallel programming.
As a result, 111 participants in hands-on training learned how to apply the POP methodology to their application codes within the first 18 months of the project. A list of training events can be found here.
Strong Emphasis on Gender Diversity and Inclusion
We strongly believe in gender diversity and inclusion, and we are committed to promoting these values in all aspects of our work. Building on our success in organising two HPC training sessions for women by women during POP2 in 2021 and 2022, we continue to consider inclusivity and diversity when organising events. When organising two workshops at the HiPEAC25 conference togther with other CoEs, we paid special attention to achieving gender balance when inviting speakers, which resulted in 74% of speakers being female. Even more encouraging was the gender balance among workshop attendees: 56% women and 44% men, resulting in a balanced gender distribution. When you consider that 74% of all conference attendees were male, this represents a remarkable achievement. Finally, POP experts co-organised the ISC25 BoF "Super(computing)heroes" and provided two speakers for the "Gender Balance in HPC" panel at the EuroHPC Summit.
World-renowned Tools
The performance analysis, energy-efficiency, and correctness tools created and developed by the POP project partners BSC, JSC, RWTH, UVSQ, INESC-ID and IT4I provide the basis for our insightful performance assessment services. All tools got enhanced to collect and present the common POP metrics, which we developed during the course of the POP projects. In additional, many tool enhancements have been designed and implemented to enable the use of the tools on all current EuroHPC computing platforms and the latest versions of the parallel programming APIs.
This includes providing Hybrid MPI+GPU Metrics in TALP, delivering an easy and extremely portable access to CUBE for performance analysts, allowing to analyse HIP codes with Extrae and Paraver, an integration of the CARM and Paraver tools, and automation of the calculation of the POP metrics with with cube_pop_metrics and POPAdvisor.
Scientific Impact via Peer-reviewed Publications
As POP3 is primarily a service-oriented CoE, we are very proud to still be able to publish scientific papers in peer-reviewed journals and conference proceedings. Topics published fall mainly into two categories: papers about enhancements to POP tools, methodology or metrics, and joint publications with customers about successful collaborations and performance assessments.
Examples of such papers published during POP3 include an summary of 15+ years of joint parallel application performance analysis/tools training with Scalasca/Score-P and Paraver/Extrae toolsets, a Methodology for GPU Frequency Switching Latency Measurement, an Exploration of Domain-Specific Features for Epistasis Detection Acceleration on Tensor Cores called EPIClear, towards Exascale Computing for Astrophysical Simulation Leveraging the Leonardo EuroHPC System, the Performance Analysis of Vlasiator at Production-Scale, a method for Tracking the Critical Path of Execution for GPU Offloading Applications, an in-depth study of GPU frequency-scaling latency and its optimization on modern architectures, and finally Leveraging NVML GPM for NVIDIA GPU Monitoring.
Older Sucess Stories and Customer Testimonials
from POP1 and POP2 can be found here.