Efficient help for supercomputing


Consulting and supporting at coding: The mentors at LRZ help users at supercomputing and with algorithms. Photo: J. Schnobrich/Unsplash

Radar satellites take images of the Earth even when it is cloudy. To do this, they send their signals to Earth and receive their echoes. Complex processes produce images that compose structures on the surface from millions of dots. They make roads recognizable, green areas, even buildings of different heights. SuperMUC-NG creates three-dimensional city and building models from masses of these images, the results will be used to research urbanization by at a team led by data scientists Professor Xiaoxiang Zhu and Dr. Yuanyuan Wang at the Technical University of Munich (TUM) and the German Aerospace Center (DLR). "The biggest challenge in our project is storing petabytes of data," Wang explains.

Support for high-performance computing, or HPC, projects has been improved considerably through a mentoring program run by the Gauss Centre for Supercomputing (GCS), of which the LRZ is a member. "Computing time and supercomputing resources are valuable, and we want them to be used efficiently," explains Dr. Gerald Matthias, who heads the Computational X Support group at Leibniz Supercomputing Centre (LRZ). "That's why we've been intensively supporting computational projects since the end of 2019, and we provide researchers with a mentor." About half of the computational projects at the SuperMUC-NG, which may well take years to complete, involve questions about software and codes or how to handle huge data sets. Addressing all 6480 compute nodes of the SuperMUC-NG with one algorithm is an art in itself: "For users, the focus is on simulations and data, not necessarily computer technology," says Matthias. "They are, after all, primarily scientists who develop applications for their projects."

Optimize, advise, ask

Wang needed advice on data storage: The radar images are filtered before being analyzed, automatically optimized, and supplemented with additional information if necessary in order to compensate for blurring or signal-to-noise ratios. Faced with this task, even supercomputers quickly reach their limits. The mentor of the project at the LRZ, seismologist André Kurzmann, PhD, opened the way to the new Data Science Storage system for outsourcing image and simulation data. The mentors also help with the optimization and implementation of algorithms and bring researchers in contact with other LRZ teams, for example when researchers need artificial intelligence methods, when it comes to the cataloging and searchability of research results, or when simulations need to be visualized: "As a mentor, I look at how the project is progressing and follow-up if nothing has happened for a long time," Kurzmann reports. "Many of our users are hesitant to approach us themselves. The danger is great that computing time is then not used and projects are canceled."

Kurzmann and his 10 colleagues are involved in the projects from the beginning, they can recommend HPC software for the planned computations or strategies for processing unstructured data, provide support with technical questions or clarify details about the computing time request and its extension. In addition to astrophysics, medicine, natural sciences, and engineering, diverse disciplines are represented on the team; where possible, mentor and mentee are matched based on domain knowledge. Zhu and Wang's project started in 2017; in the meantime, the first 3D city models are available and are being processed and researched in more detail. "Having a direct contact person at the LRZ allows us to address problems more quickly to the person who will provide a solution," says Wang. A side effect of mentoring: contacts deepen, new connections are created. "Through André Kurzmann, we have established new collaborations with the LRZ; he is also active as a scientific advisor in our ERC project So2Sat and participates in the annual project meeting."

"Short reaction time boosts the project"

Physicist Prof. Dr. Francesco Knechtli and his team at the University of Wuppertal work on quantum chromodynamics, or the assembly of hadrons from elementary particles, and calculated their energy spectrum at the SuperMUC-NG.

What did you calculate or model? Prof. Dr. Francesco Knechtli: Our project deals with hadrons consisting of a charm-quark-anti-quark pair, the charmonium. Whereas pair annihilation was previously neglected in calculations of the energy spectrum, we calculate these contributions directly by optimizing the technique of distillation for the quark fields.

How did the HPC team support you in this? Knechtli: We analyze tens of thousands of configurations of the gluon fields. To calculate the various quantum numbers, we need an immense amount of RAM. André Kurzmann also helped us handle the simulations.

What technical problems were you able to solve? Knechtli: Smaller simulations were bundled to increase the throughput of larger jobs. More RAM allowed us to take full advantage of distillation. The project gained access to the Data Science Archive, and we were able to copy configurations from the JUWELS supercomputer in Jülich to the SuperMUC-NG.

Does mentoring push the projects? How? Knechtli: The short response time to questions has helped the project.

How far has your project progressed? Knechtli: Pair annihilation of charm quarks and anti-quarks contributes to the mass of the charmonium. With our calculations, we succeeded in directly calculating this effect for the first time, albeit with charm quarks half as heavy. Now we are optimizing the numerical algorithms to be able to calculate the charmonium as in nature in the future.


Prof. Dr. Francesco Knechtl, University Wuppertal