CFD Online Discussion Forums

CFD Online Discussion Forums (https://www.cfd-online.com/Forums/)
-   CFX (https://www.cfd-online.com/Forums/cfx/)
-   -   Problem with CFX solution speed (https://www.cfd-online.com/Forums/cfx/149669-problem-cfx-solution-speed.html)

reza_k March 6, 2015 06:39

Problem with CFX solution speed
 
Dear experts,
I am a beginner at ANSYS CFX, and I have some problems with solution speed.

I am working on numerical simulation of film cooling. The domain Contains three zones,

1- Main (which could be assumed as a long flat plate)
2- The Plenum (contains coolant fluid)
3- The Passage (which is the link between plenum and the main).

Solution is conducted by:

a. CPU : Core i7-4820k
b. Parallel solver, 4 Real cores (just about 2 times faster than serial !?)
c. 3.2 million meshes
d. Two translational periodic interfaces exist in the domain.
e. average y+ = 0.5
f. material: Air Ideal gas
g. Turbulence model: SST-Kw
h. Solution is Steady state


According to these settings, each iteration takes about 100sec. I know CFX is fully a coupled solver and consequently
solution takes more time than a segregated... I would appreciate if some experts help find the answers of the bellows questions:


Q.1 - How much time is reasonable for each iteration for these kind of problems.

Q.2 - Is there any way to solve this problem faster without changing the hardware?


Thanks a lot...

monkey1 March 6, 2015 07:45

I can only give you one hint concerning computing speed. According to a sentence I once heard from ANSYS support, the optimum for CFX is: 250.000 Elements / core. Meaning for a 3,2 Mio Element Mesh you would need a least 3,2*4=12 (13) cores.
And apart from reducing ne number of elements, or the number of solved equations (meaning if you only solve the flow without heat transfer or any extra equations), or reduce the numerical quality (by increasing the convergence criteria, switching the solver from double to single precision, avoiding High resolution solving schemes, etc...) I would not have a tip how to increase computing speed.
Except for one:
Carefully check your mesh quality. The less "bad" elements (Quality criterion in ICEM should be >0,2 for Tetraeders and >0,1 for Prism layers) you have, the faster it converges and solves. Also from my experience, a structured mesh gives a faster solution than an unstructured.

ghorrocks March 6, 2015 07:48

Your mesh is 3.2m nodes/elements. Have you shown you need a mesh that fine? You might have made it finer than is required. Alternately your mesh might not be fine enough and you need it finer. Unless you have checked you do not know and you are just guessing - and unless you check you will have no idea whether your results are accurate or not.

So generate another mesh with half the element edge length (so approx 1/8 the number of nodes) and compare the fully converged results. If there is no difference then the coarser mesh is fine.

And if it says you need the finer mesh then you will just have to put up with long run times. The longest simulation I have done ran for 6 weeks. You would be able to do 36288 iterations in that time - that should be enough to get convergence :).

reza_k March 6, 2015 09:34

Thanks monkey1, thanks ghorrocks...
I really appreciate your guidance.

Unfortunately, I have checked most of the points you mentioned. The 3.2m elements have been generated for Wall Modeled Large Eddy Simulation and the current SST simulation is just an initial condition. However based on the literature at least 5m elements are needed. So it doesn't sound good, if there is no way to increase the solution speed rather than improving hardwares.

ghorrocks March 6, 2015 21:04

CFD is the just about the most computational intensive science about. And if you are doing LES then you have to expect long run time. That's why CFD uses a lot of parallelisation - so look at putting more computers on your network to speed things up.


All times are GMT -4. The time now is 03:10.