2
$\begingroup$

Using an established criterion involving capturing eddies down to the Kolmogorov length scale it can be reasoned that the order of grid points in the computational mesh needs to be $N^3 \ge Re^{9/4}$ where Re is the Reynolds number of the flow. If $Re \approx 50,000$ then I get about $N^3 \approx 3.7\times 10^{10}$. I know very little about CFD but I am wondering if this is a outrageous requirement given today's high performance computers?

I can get access to nodes of a supercomputer cluster. Im not sure of the exact specifications of the cluster, nor which specs are relevant for our purposes of calculating the total time, so if someone could make an assumption and base their calculations on this spec, I could modify my expected computing time once I know what im dealing with.

Improve this question
$\endgroup$
3
  • $\begingroup$ What specifically do you want to simulate? $\endgroup$
    – Rhys Ulerich
    Nov 29, 2014 at 21:38
  • $\begingroup$ Thanks for your response- I want to simulate flow over an airfoil or if possible a full 3D compressor blade in an annulus. $\endgroup$
    – Dipole
    Dec 1, 2014 at 21:36
  • $\begingroup$ As Bill commented, generally this would be done using RANS or LES on account of the expense inherent in tackling such a geometry. Nothing says you can't directly simulate it aside from the cost, however, so if you can get the allocation and believe you can provide good inflow and outflow treatment, Godspeed. $\endgroup$
    – Rhys Ulerich
    Dec 2, 2014 at 2:24

2 Answers 2

Reset to default
3
$\begingroup$

$3.7\cdot 10^{10}$ is feasible today on the very largest of computers, but barely so. Recall that you also have to do a lot of time steps (with shorter time step size the larger the Reynolds number is), so it is important that the local problems on every processor remain small. A reasonable number of unknowns per processor may be 100,000 which would translate to needing 370,000 processors -- possible today, but definitely not easily.

Improve this answer
$\endgroup$
3
$\begingroup$

We have people doing $4096^3$ and $8192^3$ mesh points 3D domains on our systems at TACC using 16384 cores (1024 nodes). These simulations take months of running time to achieve the results the scientists want.

Your Reynolds number is equivalent to a $\sim 3332^3$ domain, which is smaller than what we have running on our systems, but not by much. You're going to need a large computer, but not, by a fair amount, the largest in the world.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks ill have to get back to you once I know what I can get a hold of. $\endgroup$
    – Dipole
    Dec 1, 2014 at 21:38
  • $\begingroup$ I should say, I was talking about isotropic, free-air turbulence and flows in channels. Flows over complex domains are not often done with DNS solvers. My experience is that RANS and LES codes are more common for problems with interesting geometries. $\endgroup$
    – Bill Barth
    Dec 1, 2014 at 22:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.