Modular High-Performance Computing

Modular HPC is an approach that combines different kinds of processing, memory and accelerators into a single heterogeneous system that can provide just the right mix of resources for a given task. Due to its modular nature, such systems can be extended with novel hardware like memristors, neuromorphic computing or even quantum compute resources. Modular HPC systems can accelerate tasks like genomics data processing in PRECISE, time series data captured from animal models or imaging data from microscopy or MRI. We were the first group outside of HPE to work on Memory-Driven Computing and continue this work with the memory-centric compute resources at the DZNE. One major field of work is to develop tools that can be easily integrated to accelerate existing application and enable new research directions.

Distributed infrastructure and tools

Large-scale studies like DELCODE show that research needs to scale in size and geographical distribution. The compute resources need to follow that and centralized architectures are becoming less and less feasible with massively growing data set sizes. Scientific computing infrastructures and tools need to reflect this. We investigate how distributed systems can bring together sites that produce large data sets like the genomics research in Bonn, Göttingen, and Tübingen or MR imaging in Magdeburg and Bonn. Besides such a service mesh, we also apply Swarm Learning to new domains and help privacy-preserving collaboration.

Patient data security and privacy in sharing and collaboration

Researchers collaborates with partners all over the world and relies on sharing of large and sensitive patient data sets. Recent regulatory developments like GDPR introduce uncertainty and overheads, here novel privacy preserving approaches are needed that rely on synthesized data or machine learning techniques like swarm learning. In the project Pro-Gene-Gen we develop such tools that protect patient data in collaboration with CISPA.