Software performance
In many areas of research today — whether it’s Magnetohydrodynamics, Quantum Mechanics, or Fluid Dynamics — software is at the heart of how we analyze data and develop new insights. The performance of this software matters more than you might think. It not only cuts down on the waiting time for simulations to complete, but it also unlocks new avenues of research that were once limited by computational constraints, allowing for finer grids, smaller timesteps and more accurate physics. It’s not just about saving time, but also about enabling better science.
How can we help you?
- Profiling and Benchmarking: Using advanced tools like TAU, HPCToolkit, AMDuProf, and Intel VTune—alongside custom-built utilities—we can analyze performance issues such as cache misses, memory inefficiencies, pipeline stalls, and parallelization bottlenecks. This in-depth profiling allows for highly targeted optimizations.
- Parallelization: Be it restructuring serial code for parallel execution or fine-tuning existing parallel implementations. We have extensive experience in optimizing algorithms for distributed computing environments using MPI and OpenMP.
- Hardware acceleration: We specialize in adapting algorithms to harness the power of GPUs using CUDA, KOKKOS, or Alpaka. Additionally, by leveraging vector extensions like AVX on modern CPUs, we are able to perform multiple computations per instruction, delivering substantial performance gains.
- Code refactoring: Legacy codebases can become difficult to maintain and underperform on current-generation hardware. We have a deep understanding of older languages like Fortran77, Fortran90, and C90, and can refactor outdated code to improve performance, readability, and maintainability.
