Posters at SciCoCo.nrw: Training LLMs and an Integrated Approach for Benchmarking
Caro's Poster: Best Practices for Training LLMs on HPC Systems: https://juser.fz-juelich.de/record/1061381
Mike's Poster: A Benchmarker's Journey: An Integrated Approach from Onboarding to Benchmark Visualization: https://juser.fz-juelich.de/record/1061382
From September 21 to 23, 2026, SciCoCo.nrw – Science, Compute, Connect took place in Aachen. The hands-on community event gathered researchers, software engineers, data stewards and AI enthusiasts from throughout the region. Workshops offered the chance to gain new skills, and Birds-of-a-Feather (BoF) sessions provided a platform for open discussions on interesting topics. In keynotes from the fields of artificial intelligence, research software engineering, high performance computing, and research data management, speakers shared their experiences. Thanks to Aachen’s proximity to Jülich, the Jülich Supercomputing Centre was well represented. From the Accelerating Devices Lab, Caro, Mike and Niro attended.
Photo: © Conor Crowe (not covered by this post's CC BY-SA license)
Caro presented a poster about best practices for training large language models (LLMs) on HPC systems, based on a report we shared through arXiv a while ago. It includes performance results for 3D parallelism and evaluates the impact of various features. When scaling to a large number of nodes, unique technical challenges need to be addressed to use resources efficiently, making large-scale LLM training a full-stack engineering problem.
Mike’s poster presented the software projects exaCB, JUBE and LLview, all of which are developed at the Jülich Supercomputing Centre. In concert, they form a workflow that simplifies continuous benchmarking of applications on HPC systems. The exaCB tool defines reproducible benchmark runs in the application’s Git repository. These are executed using the JUBE benchmarking framework, which expands the declared parameter space. The job monitoring tool LLview visualizes the results in a dashboard. Validators can be set up to continuously detect anomalies that could, for example, indicate a system problem, allowing support to act quickly or provide direct information about the performance regression for application developers. Both posters are embedded below.
Photo: © Conor Crowe (not covered by this post's CC BY-SA license)
In the code-along skill-up session “Empowering everyone to build reliable and efficient software with Rust”, we learned about the powerful design principles on which the Rust programming language is based. Tooling with rustup and Cargo was demonstrated, and concepts such as ownership, borrowing, aliasing and lifetimes were introduced to underline the advantages compared to, for instance, C or C++. Other sessions introduced useful tools for reproducible science and research software engineering, such as Pixi and uv.
Niro attended two workshops, “Getting started on NRW’s academic HPC infrastructure” and “Containerizing HPC Workflows with Singularity and Apptainer”. Participants in both workshops were granted access to our JURECA-DC supercomputer in order to provide interactive sessions. The first workshop introduced necessary commands to connect with the machine, to work with Linux as well as Slurm and to run the provided scripts. Participants launched GPU jobs and analyzed the submitted jobs using benchmarks. It ended with advanced topics such as HPC file systems. The second workshop focused on containers, more precisely on Apptainer. After introducing foundational topics, for example the structure of a definition file, the workshop focused on running prepared containers on JURECA-DC and ML workloads inside them.
Caro also chaired the Women in HPC community meetup, an engaging session that brought together the WHPC chapters in North Rhine-Westphalia. Learn more in a blog post on the JuWinHPC voices blog.
The BoF session “Reproducible Science: Tools in Comparison” led to very insightful discussions among participants, who shared their issues and experiences regarding reproducibility. From the point of view of HPC system support, it was particularly interesting to hear about the problems that users and developers from different scientific disciplines face in making their research reproducible and hear their interpretation of what reproducibility means for them.
With its interactive formats and engaging coffee breaks, SciCoCo.nrw was a great opportunity to network and connect with partners from the HPC.nrw project, as well as with JSC colleagues from different divisions.