Cache-driven super-linear scaling in a Python fluid solver
SC26 poster and research summary on DHARA, cache-driven super-linear CPU scaling, and GPU performance at extreme scale.
SC26 · POSTER · NOVEMBER 2026
- Tandon School of Engineering, New York University, USA
- Indian Institute of Technology Kanpur, India
- King Abdullah University of Science and Technology, Saudi Arabia
When more cores do more than share the work
Simulating turbulent flows in stars and Earth's atmosphere requires resolving structures across a vast range of scales. DHARA is a Python framework for these demanding calculations, combining a readable scientific interface with computational kernels and parallel execution on CPUs and GPUs.
On KAUST's Shaheen III, DHARA scales to 786,432 CPU cores for compressible convection. Its strong scaling can be super-linear: increasing the core count reduces the time per step by more than the increase in computing resources alone would suggest. Profiling traces this effect to the memory hierarchy. Smaller subdomains fit into fast processor caches, reducing cache misses and improving per-core throughput.
The work also examines GPU scaling on Frontier at OLCF and Polaris at ALCF, showing how the same framework supports extreme-scale simulations across different computing architectures.