[JuPa]

I've read a paper which says that a transient simulation took 6 months to complete using a time step size of at least 1e-6 s on a standard quad core desktop machine (Intel i7 or equivalent).

My colleague argues that any acceleration by parallelisation in transient simulations are not possible.

I'm not sure I agree except for the fact that the minimum time a transient simulation will take is the total run time specified. I.e. if you ask CFX to solve a transient simulation for 10 hours, then it will take a minimum of 10 hours to compute using a time step of 1 hour. This assumes it takes 1 hour to compute the data in a 1 hour time step! All very confusing.

I understand there will be bottlenecks. I.e. if you compute your simulation using 100 nodes then the speed of the information transferred between those nodes may be the bottleneck.

The point of this thread/debate is this: do you foresee any issues with computing a transient simulation using parallel computing?

[ghorrocks]

My colleague argues that any acceleration by parallelisation in transient simulations are not possible.

What a load of rubbish. (You can quote me on that ). Your colleague has clearly not understood the mathematics of CFD, or done much in the way of CFD either. I can assure you parallelisation works just fine in transient simulations. In fact because transient simulations are almost always longer CPU time than steady state runs that makes parallelisation even more important than for steady state runs.

But I must say I fail to see what point you are making about total run time specified either.

There are many issues to consider when setting up parallel simulations. Small parallel simulations (less than 8 nodes or so) require no special consideration and generally work pretty well no matter what you do; medium simulations with 8-100 nodes will benefit by careful system design; and simulations or over 100 nodes will not get anywhere without a careful system design to handle the special requirements of big transient simulations.

[Lance]

And if your colleague still dont trust you, show the results from a benchmark I did. CFX 14.0, transient 30 iterations, 100 million cells, x-axis is CPU cores, y-axis is wall clock time. There was a speedup until ~900 cores, then there were too many bottlenecks for this case and the simulation became slower with more cores.

[JuPa]

Thank you all for your replies. I knew he was wrong. I just needed a sanity check.