CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   CFX (http://www.cfd-online.com/Forums/cfx/)
-   -   Memory or computing power (http://www.cfd-online.com/Forums/cfx/112723-memory-computing-power.html)

Spookz February 4, 2013 06:29

Memory or computing power
 
Hello,

I am doing a heat transfer simulation of a tool placed in an autoclave.
The heat transfer is only through convection (no radiation) and the temperature in the autoclave is slowly rising.
Real simulation time is 28 minutes and the timestep is 0.001 s
The element count of the whole mesh is 9.5 Million elements.

Actually I just want to know what a system you would recommend for this job.
A system with much memory or a system with much computing power?
What is the average time a system should work on this job?

Thanks for any answer!

ghorrocks February 4, 2013 06:52

A fast CPU is the most important for CFD. And lots of them so you can run in parallel. Memory is less important, but you need enough memory the run in RAM. If you run out of RAM it is a disaster, but if you have enough RAM to run then more RAM makes no difference. But pre- and post processing can require lots of RAM, more RAM than the solver in some cases.

I cannot estimate average time, it varies by so many things.

How do you know the time step size is 0.001s and the mesh is 9.5M elements? Did you just guess that or did you actually determine what was required?

Spookz February 4, 2013 07:08

Thanks for the answer!

I determined the element count by some test simulations and by decreasing element size. With a timestep of 0.001 I got convergence.
Would you say a system with 24 CPU-cores (each 2.9 Ghz) and 10 GB RAM is enough?

I need to finish this simulation in less than a month :)

ghorrocks February 4, 2013 17:25

So did you do a sensitivity analysis to determine the mesh size? It sounds like you did not do a sensitivity analysis to set the time step size. This means you will spend the next month finishing your simulation and will get a very inaccurate answer. Does not sound very constructive to me.

You need to do a sensitivity study on mesh size, time step size, convergence tolerance and any other tunable parameter first. Once you know all these parameters then you know how big the simulation will be and then you can arrange for a computer big enough to solve it in the time you have available.

If you do this backwards you will get a simulation finished on time whose results are meaningless.


All times are GMT -4. The time now is 15:32.