Hello,
I have a question about the maximum number of processor for parallel calculation.
I use the version 2RC1 and with the GUI I can't select more than 32 procs.
If I put 1 proc in the GUI and ask in the runcase 40 procs, the simulation starts and seems ok ,I obtained the file domain_number_40.
Is this simulation good? Is there a problem of communication with a big number of procs?
Thank you for your answer.
Sylvain
maximum number of processors
Forum rules
Please read the forum usage recommendations before posting.
Please read the forum usage recommendations before posting.
-
Mickael Hassanaly
Re: maximum number of processors
Hello Sylvain,
The runcase takes the priority over the GUI. It's more a question of processus than processors. So your simulation should be ok.
Mickaël
The runcase takes the priority over the GUI. It's more a question of processus than processors. So your simulation should be ok.
Mickaël
-
Yvan Fournier
Re: maximum number of processors
Hello,
The limitation in the GUI is a bug (a maximum value is enforced to avoid "unrealistic" values, but it is much too low for practical purposes). For big cases, we usually edit the runcase outside the GUI, so we did not notice it.
I'll fix the bug for future releases (all that is required is to replace 32 by some higher value in gui/Pages/BatchRunningForm.ui line 247, and reinstall the code; setting the box width to 100 or 150 instead of 50 in the same file line 239 also helps).
Depending on the machine, as little as 3000 (for Blue Gene type machines) to 10000 (for classical clusters) cells/rank is reasonable, so for a 5 million cell mesh, you could expect reasonable speed-up up to 5000000/10000 = 500 processors (for larger processor counts, it may be a little lower, so you'll see little gain between 256 and 512 processors for a 10 million cell mesh, but this rules gives you an idea of what is realistic).
The highest number of cores Code_Saturne has been tested on is 32768, so the 32 process limit in the GUI is a bit frustrating...
Best regards,
Yvan
The limitation in the GUI is a bug (a maximum value is enforced to avoid "unrealistic" values, but it is much too low for practical purposes). For big cases, we usually edit the runcase outside the GUI, so we did not notice it.
I'll fix the bug for future releases (all that is required is to replace 32 by some higher value in gui/Pages/BatchRunningForm.ui line 247, and reinstall the code; setting the box width to 100 or 150 instead of 50 in the same file line 239 also helps).
Depending on the machine, as little as 3000 (for Blue Gene type machines) to 10000 (for classical clusters) cells/rank is reasonable, so for a 5 million cell mesh, you could expect reasonable speed-up up to 5000000/10000 = 500 processors (for larger processor counts, it may be a little lower, so you'll see little gain between 256 and 512 processors for a 10 million cell mesh, but this rules gives you an idea of what is realistic).
The highest number of cores Code_Saturne has been tested on is 32768, so the 32 process limit in the GUI is a bit frustrating...
Best regards,
Yvan
-
Serra Sylvain
Re: maximum number of processors
Hello
I only want to thank you for your answers.
(it should interesting to add a flag that correspond to "finished conversation" on the forum…)
Best regards,
Sylvain
I only want to thank you for your answers.
(it should interesting to add a flag that correspond to "finished conversation" on the forum…)
Best regards,
Sylvain