Project: Supercomputer Research and Performance Optimization
A supercomputer does not run fast simply because it has been installed. We work on both sides of the problem — the design of the machines and the programs that run on them — to draw out the performance they are capable of.
What we work on
Our targets span architectures of very different character: vector supercomputers, GPUs, and emerging instruction sets. On these machines we run both real applications — disaster prevention, turbulence, fluid analysis, machine learning — and standard benchmarks that measure a machine's raw capability, to find out where performance hits its ceiling. What we learn feeds back into how the programs are written and how the data is laid out.
Why it matters
Computers rarely deliver the performance their hardware specifications promise. Depending on how a program is written and how its data is arranged, the same calculation can run several times slower. Closing that gap matters directly in fields where computing time shapes the outcome, from tsunami damage estimation to product design.
Main activities
- Performance evaluation of the vector supercomputer SX-Aurora TSUBASA (reported at SC18)
- Accelerating tsunami inundation simulation and applying it to real-time damage estimation
- Optimizing direct numerical simulation (DNS) codes for turbulence, with I/O performance evaluation including on the supercomputer Fugaku
- GPU parallelization of the BCM fluid-analysis method, and optimization of steam-turbine simulations
- Performance analysis with standard benchmarks such as the Himeno benchmark and HPCG
- Graph processing and sparse-matrix operations implemented for vector machines
- Automating optimization with the Xevolver code-transformation framework
Keywords
High-performance computing (HPC) / Vector architectures / GPUs / Benchmarking / Performance optimization