TGV performance

I am wondering if someone can comment on the TGV results (case2) in Loppi et al. Computer Physics Communications 233 (2018). It looks like the wall-clock time is quite high compared to incompressible flow solvers. Is this because of the method to handle low Mach number flows with a compressible code?

It is difficult to answer as you are not giving any details. What are you comparing against (resolution, hardware, wall-time, discretisation, time-integration etc…)?

I think our 3h38min wall-time is fine. For example here https://arxiv.org/pdf/1802.01439.pdf similar setup with an incompressible solver took 4h.

Here are few things to consider:

  • The number of elements is 52^3 but the number of solution points is 260^3 (P=4 polynomials)
  • Since the geometry is just a box, purely spectral / high-order finite difference schemes have an advantage.
  • You could strong scale. Double the number of GPUs and cut the wall-time in half.
  • We have made the solver faster since by adding other convergence acceleration techniques. Latest release should be ~2-3x faster.

Niki