CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   FLUENT (http://www.cfd-online.com/Forums/fluent/)
-   -   Selecting cores independently in a dual quad core machine (http://www.cfd-online.com/Forums/fluent/67999-selecting-cores-independently-dual-quad-core-machine.html)

jalarron September 2, 2009 22:45

Selecting cores independently in a dual quad core machine
 
Hi all,

I need to select specifically the cores in a dual quad core Xeon machine, running fluent in SUSE SLED 10 Sp2.

I have 8 cores, and wanted to run 2 cases each using 4 cores. I've started one case with: fluent 3d -t4 & ;but as soon as I start the other case: fluent 3d -t4 &. The first case slows down immediately. So I'm assuming that the second command is picking the same or some of the cores of the first case.

Is there a way to select independently each core in fluent? I used to do this with the option cnf=n1,n2,n3,n4. But in that time I had each node as separated pc.

Cheers,
Javier

mystic_cfd September 7, 2009 17:06

Hi Javier,

Are you running 6.3.26? There was a problem that it would launch the second run on the same cores with that version of Fluent. If you use 6.3.35 or 12.0.16 it should not happen.

To set the cores manually:
- Start the first job to run (it will grab four cores for itself).
- start second Fluent session and load case and data
- set the process affinity with the TUI command:
(set-affinity "0 4 1 5 2 6 3 7")
This example will set a 4-partition parallel run to run on 4 cores; the data is arranged in pairs such that partition 0 is sent to core 4, partition 1 is sent to core 5 etc.

To see which cores are currently loaded, use top, and when top is running press "1". This expands the CPU section at the top to show individual cores. From the load status you will be able to see which cores are not occupied.

Hopes this helps.

cheers.

PS. if you are using older series Xeon (e.g. 52xx, 53xx, 54xx, 73xx) the bad news is that there will be very little speedup going from 4 to 8 cores due to memory bandwidth limitations. for newer series (55xx) this will not be the case. if licenses are a problem, it is best to run 2 threads per CPU (only use 2 of the 4 cores). It should be as fast as using all 4, and it uses half the licenses.


All times are GMT -4. The time now is 14:29.