[peter.pan]

Hi,

I want to know for two cases using the same conditions, down to the RAM/per nodes, will using a parallel scheme improve the capture of the physics?

Suppose in Case 1 involves a million nodes and we run it serially using a 16 GB RAM system, now in Case 2 we run it in parallel using a four processor system each having 4GB RAM in such a way that the RAM/processor in both cases be same, 1 million nodes/16 GB RAM and 0.25 million nodes/4 GB RAM.

Which one would supposedly give me better results in terms of capturing physics better, ignoring the frailities of using a parallel code including ignoring Amdahl's Law.

[som87]

in short
improve computational potential==>yes
improve result/physics===>no

[peter.pan]

Thanks Somdeb for your thoughts. But I thought a better computational resolution or better computational potential should yield better results? Or better computational results also depend on factors like more number of node points etc?

[som87]

Let me elaborate on :

Originally Posted by peter.pan

I thought a better computational resolution or better computational potential should yield better results? Or better computational results also depend on factors like more number of node points etc?

Yes, better computational resolution indeed yields better result. But computational resolution is not something which depends on our computational power.

Say you are doing DNS of 3d lid driven cavity with 200x200x200 mesh.
How resolve our computation'll be, depends alot on the mesh i.e. at how many nodes you have the solution (the discrete one).

Now returning to your point, with more computational power (i.e. parallel) we can deal with more number of mesh which'll eventually provide a higher level of "resolution" in our result. But the "better" result is due to more "refined discretization" not 'cause of your 'computational power'


-----the end from my part-----

[peter.pan]

Thanks Somdeb, that makes perfect sense. Appreciate your reply