Problem

A legacy product for multiple regression had been implemented using Spark RDDs. The person who wrote the code had left. The code was becoming difficult to maintain or extend and wasteful for small datasets.

Contribution

• refactored the code, allowing for a change of computational backends
• added two more computational backends, one with PySpark and one with Pandas
• in-software switching among all three backends
• introduced tests and best software engineering practices

Outcomes

• 20x speed-up for small datasets, when the data small enough for Pandas
• 1-5x speed-up for larger datasets
• flexible and future-proof code.

Problem

client from financial sector had Excel and SQL workflows requiring a significant amount of manual effort and judgement.

Contribution

• implementation of low-code solution in client’s environment using KNIME workflows
• wrote Python code where KNIME was unable to handle complexity
• called external API for additional automation and validation
• negotiated contract with external API provider on behalf of client

Outcome

• automation saved 30 person-hours per month for client
• external API validation provided higher data quality
• project payback in < 1 year

Problem

Pre-operative planning of treatment of cardiovascular disease requires patient-individual blood flow simulations.

Contribution

• set up MPI-parallel CFD simulations on complex geometries via the Lattice-Boltzmann method
• validation of simulation results against experiment data
• set up workflow for geometry extraction, setting of boundary conditions, discretization and running of LBM simulations

Publication

Sebastian Meier, Anja Hennemuth, Nikola Tchipev, Andreas Harloff, Michael Markl, and Tobias Preußer. Towards Patient-Individual Blood Flow Simulations based on PC-MRI Measurements. In Lecture Notes in Informatics, volume P-192, page 447, Bonn, Germany, 2011

Official link: Gesellschaft für Informatik; available from: tu-berlin.de

Problem

Aircraft landing on the same runway face risk of flying into wake vortex of previous aircraft.

Contribution

• performed 27 wall-resolved large eddy CFD simulations
• simulations on SuperMUC, 512 cores each
• analyzed >3 TBytes data
• legacy Fortran modelling code translated to C++

Outcomes

• light crosswind most hazardous at airports (0.5 w0)
• mechanism of ascent, descent, rebound and decay explained
• C++ code used for real-time vortex prediction onboard aircraft

Publication

Frank Holzäpfel, Nikola Tchipev, and Anton Stephan. Wind Impact on Single Vortices and Counterrotating Vortex Pairs in Ground Proximity. Flow, Turbulence and Combustion, 97(3):829–848, 2016

Official link: springer.com; available from scispace.com.

Problem

Routines for compaction or merging of shapes in transistor layout tools can be O(N²) computationally expensive, which can lead to hours of waiting time for users.

Contribution

• optimization flags improvement: 5x - 20x faster
• look-up table bottleneck: 3x - 10x faster
• multi-threading on 16 threads: up to 10x faster
• legacy code debugging, testing and improvements

Outcomes

• compaction: 500x faster
• merging: 130x faster
• user waiting time cut from 8+ hours to 5 minutes

Problem

One mililiter of water (1 mL) contains more than 10¹⁰ H₂O molecules. How many molecules can be simulated on the largest supercomputers today?

Contributions

• Introduced
  • OpenMP in the whole code base
  • SIMD wrappers
  • overlapping of MPI collectives
• 27% more molecules
• 43% faster weak scaling
• 146% faster strong scaling
• 256-bit SIMD vectorization

Outcomes

• First simulation of 2 × 10¹⁰ molecules
• Total of 344,064 threads in parallel
• 1.3 PFLOP/sec, 0.67 PBytes RAM
• Award from the President of Bulgaria

Publication

Nikola Tchipev, Steffen Seckler, Matthias Heinen, Jadran Vrabec, Fabio Gratl, Martin Horsch, Martin Bernreuther, Colin W. Glass, Christoph Niethammer, Nicolay Hammer, Bernd Krischok, Michael Resch, Dieter Kranzlmüller, Hans Hasse, Hans-Joachim Bungartz, and Philipp Neumann. TweTriS: Twenty Trillion-atom Simulation. The International Journal of High Performance Computing Applications, 33(5):838–854, 2019.

Official link: sagepub.com. Available from researchgate.net.