Page 1 of 2

Breakdown on cores?

PostPosted: Fri Sep 06, 2013 6:00 pm
by rickoic
Ok, have 4 x 12 core AMD 2.5GHz processors
Have slots for 3 x 16 GPU's (Installed eVGA GTX 660's)
Dropped core count on my bigadv from 48 to 45 as breakdow for;
48 = 2 2 2 2 3
45 = 3 3 5
40 = 2 2 2 5

Now all breakouts for 45 are prime numbers but they are considered small primes.

When bigadv restarted it gave me a line

Mapping NT from 45 to 40

Is it because 45 gives all primes that it dropped it all the way to 40?

Tks
Rick

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 6:14 pm
by Napoleon
Does it drop all the way to 40 from 47 and 46 as well?

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 6:31 pm
by Joe_H
I think past a certain point the heuristic used will kick out odd numbers. Then 44 would get bounced as a multiple of 11, 43 & 41 are large primes, and the even number 42 is a multiple of 7 which sometimes is problematical. Multiples of 5 usually are okay, but recently one project was documented to not work well with a SMP setting of 5 or its multiples.

As I understand it, the Gromacs code used should work with almost any SMP setting. In practice PG found various settings that had higher failure rates or would not work at all. The rule of thumb as expressed here comes down to "large" primes and odd numbers can cause issues. "Large" was never defined to be an exact value, but any prime 11 or larger appears to be excluded. On a project by project basis 7 is sometimes excluded as a usable SMP setting. On the low end there are some projects limited to a maximum setting of 2 as they involve a small number of atoms.

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 7:54 pm
by rickoic
Napoleon wrote:Does it drop all the way to 40 from 47 and 46 as well?

Yes it does.

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 8:04 pm
by Joe_H
As expected. 47 is prime and 46 is a multiple of another large prime - 23.

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 10:59 pm
by bruce
rickoic wrote:48 = 2 2 2 2 3
45 = 3 3 5
40 = 2 2 2 5


GROMACS prints the actual breakdown that it uses in it's log, though that's not typically displayed where we see it. The choices that are excluded is a FAH add-on, not part of native GROMACS.

I think it's always three dimensional so 48 might be 3 4 4 and 40 might be 5 4 2. I don't have access to the real data, so somebody that does can correct me.

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 11:18 pm
by 7im
Decomp is always in 3 dimensions, but not always 3 dimensional.

See PantherX's post: http://foldingforum.org/viewtopic.php?f=66&t=22884&p=242642

So not only is there a relationship to how it breaks down in to 3 parts but also, if the WU itself supports 1D, 2D, 3D, etc.

Possible examples...
48 cores in 3D is 3 4 4
48 cores in 2D is 2 3 8.
48 cores in 1D is 1 2 24, or 1 4 12, if that's even supported.

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 11:33 pm
by rickoic
7im wrote:Decomp is always in 3 dimensions, but not always 3 dimensional.

So not only is there a relationship to how it breaks down in to 3 parts but also, if the WU itself supports 1D, 2D, 3D, etc.

Possible examples...
48 cores in 3D is 3 4 4
48 cores in 2D is 2 3 8.
48 cores in 1D is 1 2 24, or 1 4 12, if that's even supported.


So; 45 cores would be 3 3 5 for all 3 as thats the only way it breaks down. But FAH doesn't like it, why? Because all 3 of the numbers are prime and theres no even number in the bunch?

Re: Breakdown on cores?

PostPosted: Fri Sep 06, 2013 11:46 pm
by 7im
That project may not support 3D decomp. It may only support 2D decomp, or 1D and 2D.

Found these log examples... so I'm not sure what selects 1D vs. 2D vs. 3D decomp.

Code: Select all
[10:38:45] Mapping NT from 48 to 48
Starting 48 threads
Making 2D domain decomposition 8 x 6 x 1
starting mdrun 'Overlay'
[10:38:50] Completed 0 out of 250000 steps (0%)


A 32 core box gave this message: Making 3D domain decomposition 4 x 4 x 2

Re: Breakdown on cores?

PostPosted: Sat Sep 07, 2013 3:12 am
by rickoic
First wu was P8101 (Run 21, Cloine 2, Gen 271)
just finished it and got a P8104 (Run 0, Clone 75, Gen 123)
and both dropped 47 cores down to 40.

So how is the decision to do 3D, 2D, or 1D made? And is there anything the folder can do to influence that decision?

Re: Breakdown on cores?

PostPosted: Sat Sep 07, 2013 3:27 am
by PantherX
rickoic wrote:...So how is the decision to do 3D, 2D, or 1D made? And is there anything the folder can do to influence that decision?

That would be handled automatically by the FahCore and/or WU assigned. If the donor notices that their SMP:11 Slot is being remapped down to 10, then the only thing that they can do is to either change the SMP Slot from 11 to 10 to avoid the constant remapping or they can change the SMP:11 to SMP:12 so that the remapping isn't done. Do note that the remapping is always done by the FahCore and there isn't any way for it to be changed unless PG releases a newer version of FahCore.

Re: Breakdown on cores?

PostPosted: Sat Sep 07, 2013 5:48 am
by P5-133XL
As a side note: Do not ever specify more cores than you have available to fold. Using less cores than possible is far more productive than over subscribing the number of cores.

Re: Breakdown on cores?

PostPosted: Sun Sep 08, 2013 1:40 pm
by Nathan_P
P5-133XL wrote:As a side note: Do not ever specify more cores than you have available to fold. Using less cores than possible is far more productive than over subscribing the number of cores.


This, if you specify a larger number in order to use the 5 "spare" cores you will probably end up back at smp 48 and slow production right down as the BA WU and the NV driver fight for usage on the 3 cores that NV needs.

Personally I would put the GPU's into a separate machine and let that 4p run at full speed

Re: Breakdown on cores?

PostPosted: Sun Dec 15, 2013 2:50 pm
by Mike_Shaffer
On a V7.3.6, 12.04.3LTS setup with 2P 2680v2 with 40 threads, I always get "Mapping NT from 40 to 32" and it will not use all 40 threads. It also does this with v6 client. So, based on this thread, that is a function of the FahCore rather than the settings in the client?


Image

Re: Breakdown on cores?

PostPosted: Sun Dec 15, 2013 3:53 pm
by bollix47
Welcome to the folding@home support forum Mike_Shaffer.

The default maximum number of cores is 32. However, it's fairly easy to change the configuration to use all 40. For v7 open FAHControl and click on Configure, then Slots. Double click on the cpu slot. In the CPUs box you'll probably see -1 ... change that to 40 ... click on Ok and then Save.

In v6 you would change the -smp option to -smp 40.