OF Parallel Processing with Core i7 - How to Handle Hyperthreading
 

Hello all,

Does anyone have experience parallel processing with the Core i7 processor? It has 4 cores, however it appears as 8 in System Monitor due to hyperthreading of the cores. My question is this: I am running a 2-D airfoil simulation in a C-grid, which breaks up very nicely into 4 parts for parallel processing, but does not break up very nicely into 8. If I run in parallel on 4 processors I will only use half of my computer's processing power... any ideas on how I can best utilise the system's resources?

Thanks,

Dan

What kind of CPU utilization do you get with an 8 processor decomposition? I use an i7 and get pretty good efficiency with 7 or 8 threads.

Hi dancfd,

I also work with i7 processors in our cluster. You already has the clues that lead to the answer: the i7 processor has 4 physical cores, that appear as 8 virtual cores due to Hyper-Treading (HT). The processing power is not doubled when using HT, the physical cores remain the same. I have tested OpenFoam in three configurations:

1. HT deactivated - running parallel with 4 nodes
2. HT activated - running parallel with 4 nodes
3. HT activated - running parallel with 8 cores

In all configurations I am saturating the processing power of the machine, independent of what the task manager says. The performance difference is very small, with an 2-3% advantage to the configuration 1 (probably due the overhead of the HT system in configurations 2 and 3).

In short, once you get all your physical cores full, there is no advantage in splitting a physical core into 2 virtual ones, and fill them up.

From my point of view you cannot estimate your performance by means of the number of cores or nodes. It heavily depends on the case you are running in parallel. I have an i7 as well and for the latest bigger case I run - that was a steam drum by the way - 5 cores (hierarchical method) was the fastest configuration...

So if you have a huge case which needs a simulation time of around 2 or 3 days, I recommend to firstly run a little parallelisation study to see which configuration needs the shortest amount of time. Thus, you can save computational time ;)

Hello all,

Thank you for the detailed responses. I will run a parallelization study and post the results here.

Regards,

Dan

Hello all,

In case anyone is still interested, I ran a 42k cell 2D airfoil C-grid mesh on 4, 6 and 8 processors, decomposed as follows with equal weighting given to each processor:

I ran 3000 timesteps in simpleFoam, and found that the 6 cores were 27% faster than 4, and 8 cores were 23% faster than 4. I did not run any tests to verify the effect of changing the processor weighting or decomposition (e.g. 2-4-1 vs 4-2-1). For your information,

Dan