Overview of the Hackathon
This Hackathon was conducted as a pilot test run for similar such events in future. The key objective was to optimize, scale and tune the user codes to achieve better performance and/or solve larger problem sizes. The key highlight of this activity was teamwork and cross disciplinary collaboration between Domain Experts and Computer Scientists. This provided a platform for participants to learn new skills and technologies which they will be using along with their domain knowledge. Being a pilot test run, the hackathon was restricted to the Department of Aerospace Engineering (with exception of a few applications from the Department of Mechanical Engineering).
Events Dates and Place
The event was jointly organized by C-DAC and Department of Aerospace Engineering, IIT Madras from 29th July 2024 to 1st August 2024.
Participating Teams
- Team HORFID: A hybrid FD/FV high order line-based solver for compressible flows on unstructured grids
- FEST-3D: A finite volume solver for the discretized Navier-Stokes equations on block-structured grids
- Compressible Multifluid: Used for simulating various compressible multifluid applications. It is based on an unstructured finite volume method
- TPS: Used for simulating the internal flow of solid rocket motors
- CompSquare: A high order structured compressible flow (CFD) solver used to study internal and external aerodynamics
- CFD of low speed reacting flow: Solves for flow (NS), temperature and species along with radiation transfer equation
- LABELS: For simulation of incompressible flows, lattice Boltzmann method
- Unified Gas Kinetics Scheme: A finite volume solver that solves the two-dimensional BGK-approximated Boltzmann transport equation
- Flapping Dynamics: Uses Immersed Boundary Method (IBM) Solver. This IBM-based CFD solver aids in analyzing aerodynamic efficiency, energy harvesting systems and bio-inspired engineering by simulating flapping wing mechanisms of birds and insects. Other code Reduced Order Modeling using Autoencoders implemented with TensorFlow, constructs a neural network-based reduced order model of the high-fidelity CFD data
Pre-Hackathon Activities (2 weeks prior to the event in online mode)
The teams were introduced to their respective mentors. As many of the codes were focussed on OpenACC based GPU implementation, a short online training programme on OpenACC was conducted for the participants.
Following activities were carried out before the actual mini-hackathon:
- Get the code compiled and run on the target platform
- Select appropriate input test cases and setup code correctness validation mechanism
- Get the code profiled with Intel VTune profiler and identify the hotspots
Conclusion
The hackathon was successfully executed, with actively engaging the participants and extending their efforts beyond the event itself. Performance improvements were impressive, with the highest speed-up reaching 386 times, and the lowest at 1.6 times.
Future Work and Suggestions
The hackathon was successfully executed, with actively engaging the participants and extending their efforts beyond the event itself. Performance improvements were impressive, with the highest speed-up reaching 386 times, and the lowest at 1.6 times.
Suggestions for future hackathons:
- Target more user/ legacy codes on GPU clusters with emerging tools (OpenACC)
- Identify user codes which can scale on bigger clusters (20 PF) and extend necessary support (including additional system time under NSM)
- Target codes catering ‘Grand Challenge Problems’
- Conduct User’s Meets at regular intervals (say monthly) for effective engagement among both domain experts and computer scientists
- Identify codes which can be catered to quantum computing using hybrid environment