Serial vs Parallel
 

Hi,

I am totally a newbie, just like to start CFD for hobby about a month from now. I have a question about serial runs and parallel runs. Which kind of simulation require serial runs and which for parallel runs?

Well, as soon as a single core machine can't handle the simulation your are trying to run, go parallel, if you have the hardware. Three things can go wrong:

a) your memory is too small (RAM) to hold the computation
b) your CPU is too slow (takes a week, a year, 100 years...)
c) you run out of disk space (ok, this is the least likely)

Cheers!

If your problem is 3D and use any turbulence model you will definitely need to run your case on a parallel code.
For 2D steady problems I think a simple serial solver will do.

Rana

You can do a simple calculation in terms of memory:
Find out how many bytes your solution needs for (e.g.) 5 state variables. then find out how many cells/integration points you have, that should give you the minimum memory requirement.
In terms of CPU time, do some benchmarking, e.g. find out how long your code needs per timestep or such...

Once you have these rough numbers in your head, you'll be able to figure out easily if a problem is doable in the time given and on the hardware you have!

Thank you for your reply,

I have already found some issues about the hardware limitations. But my real question is, which kind of simulations are run in parallel or serial? From RanaP, 3D turbulence are done in parallel runs and 2D in serial runs. So, I like to know which other simulations are done in parallel and/or serial? Thanks in advance.

That totally depends on the number of degrees of freedom your want to solve for....it's like asking which goods will fit in my car.... of course, a pound of sugar will easily fit in your car, but 100 tons won't.... it all depends on your problem.

In general, 1d and 2d test runs for code development are run in serial, probably sth like Euler 2D gasdynamics....couette flow....academic problems.....

but beyond that? parallel!

cfd codes usually iterative process. How can we handle it in 3d with million dof problem from parallel processing????

I'm sorry, I don't understand your comment/question... I just did a computation on 125000 processors with about 10 ^9 DOF.....:eek::eek:....so trust me, it works pretty well :)

Cheers!