Hi Zach,
I was interested in running a compressible simulation with PyFR. The target Reynolds number would be 2.146e+06.
OK. Out of interest, what is the Mach number?
A p4 mesh would likely consist of around 15e+06 hexes.
OK. So this would have ~ 1.8Bn solution points.
Freddie’s spreadsheet, if I’m reading correctly, estimates a 3-D double-precision Navier-Stokes problem would require 3.7e+13 bytes or 35.2GiB of memory to run.
Do you mean 35 TB?
Again, correct me if I’m wrong, but if I recall the tilmestep should be somewhere between 1/p and 1/p^2 which would be between 0.25 and 0.0625 for a 4th order simulation.
The scaling with p is, as you suggest, somewhere between 1/p and 1/p^2. But the absolute value of the time-step will depend on how you have non-dimensionalised. It you have an inflow velocity of 1 and a characteristic length of 1 I would expect a tilmestep of ~ 1e-5 - 1e-6 at this Reynolds number, depending on mesh quality.
Ideally, I would like to run this out to steady state.
What do you mean by steady-state here, since the flow is turbulent? Do you mean run it long enough such that you can gather accurate statistics. Depending on the flow problem and the quantities you want to converge, this can mean running for 10-100s of characteristic flow passes.
AWS has p3 instances with 1, 4, and 8 NVIDIA V100 cards available each with 16GB of memory, 5120 CUDA cores, and 7TFLOPS of double-precision performance. Is this problem tractable in the latest PyFR release?
We have never done a simulation at such a high Re. But it might be tractable.
Does PyFR provide an inlet boundary condition where a total pressure and total temperature profile can be specified?
We have the sub-in-ftpttang BC, which I believe now can have a spatial dependence (someone can correct me if I am wrong).
Cheers
Peter