[hitzhwan]

Is it reasonable for the calculation velocity of Fluent decreases with the increase of cpu cores, especially when the core number grows to a degree?

[flotus1]

I interpret your issue as "running a simulation on N+1 cores takes longer than on N cores"
Correct?
That's normal. How large N is when this first occurs depends on a lot of factors. One of the largest contributors is "parallel overhead". Running a simulation in parallel usually incurs a small amount of extra work that needs to be done, which isn't necessary for serial computations. This overhead can grow with N, leading to negative scaling at some point.

[hitzhwan]

Originally Posted by flotus1

I interpret your issue as "running a simulation on N+1 cores takes longer than on N cores"
Correct?
That's normal. How large N is when this first occurs depends on a lot of factors. One of the largest contributors is "parallel overhead". Running a simulation in parallel usually incurs a small amount of extra work that needs to be done, which isn't necessary for serial computations. This overhead can grow with N, leading to negative scaling at some point.

That is right. What is the defination of "parallel overhead", how can I know which core number is best for my cases, do you have any methods to predict before simulation? Thank you so much.

[hitzhwan]

Originally Posted by flotus1

I interpret your issue as "running a simulation on N+1 cores takes longer than on N cores"
Correct?
That's normal. How large N is when this first occurs depends on a lot of factors. One of the largest contributors is "parallel overhead". Running a simulation in parallel usually incurs a small amount of extra work that needs to be done, which isn't necessary for serial computations. This overhead can grow with N, leading to negative scaling at some point.

Could you tell me what is the parallel overhead? How can know which core number is suitable before simulation? Thanks so much.

[flotus1]

Could you tell me what is the parallel overhead?

I did that already. "Running a simulation in parallel usually incurs a small amount of extra work that needs to be done, which isn't necessary for serial computations"
What exactly that is depends on the implementation.

How can know which core number is suitable before simulation?

I already hinted at the fact that parallel overhead is not the only factor that can contribute to parallel efficiency decreasing at high core counts.
That's why it is hard to tell how many cores are ideal beforehand. Cell count often plays a role, but there are just too many other factors involved, even including the hardware used to run the simulation. It's trial&error before slowly building up experience with your particular simulations and hardware setup.

[hitzhwan]

Originally Posted by flotus1

I did that already. "Running a simulation in parallel usually incurs a small amount of extra work that needs to be done, which isn't necessary for serial computations"
What exactly that is depends on the implementation.


I already hinted at the fact that parallel overhead is not the only factor that can contribute to parallel efficiency decreasing at high core counts.
That's why it is hard to tell how many cores are ideal beforehand. Cell count often plays a role, but there are just too many other factors involved, even including the hardware used to run the simulation. It's trial&error before slowly building up experience with your particular simulations and hardware setup.

Thank you, if we have a same computer hardware, does the linux system have faster velocity than the windows system? What is the difference between them? Another question is there any difference of calculation velocity between the Brand machine and assembly machine？