Modify memory settings for Numforce

[Glxblt76]

Hello,

I'm using NumForce for my frequencies calculations in parallel. I use, for each calculation, 16 nodes on my CPU. I have 32 GB of RAM. However, it seems that the RAM is not completely used. The computations are still incredibly slow, even if the NumForce.out file do indicate that I use my 16 nodes for each calculation.

My question is: where to check how much memory is used per node by NumForce and how to increase it if it's too low?

Thanks in advance! 

[antti_karttunen]

Hi,

Numerical frequency calculations can be rather slow, especially for larger systems. You did not mention which level of theory you are using. Using more RAM will probably not result in any significant speedups, unless you are calculating frequencies with electron-correlated methods (ricc2 module). In this case, you should increase the $maxcor keyword to something like 1500 (in MB).

Best,
Antti

[Glxblt76]

Dear antti_karttunen,

Thanks for your help.

I'm using BP86/TZVP level and I used the COSMO solvation model during the optimization. Because of COSMO solvation model, I needed Numforce to perform my frequencies calculations and check if my structure is a true minimum.

The molecules I study have 30-140 atoms, most of them having 45-80 atoms.

My system specifications are:
Processor: Intel® Xeon(R) CPU E5-2687W 0 @ 3.10GHz × 16
RAM: 15,6 GiB (not 32 I was mistaking).
Graphics: Quadro K4000/PCIe/SSE2 (never know, if I can use the GPU to speed up a bit the computations...)

So, in your opinion, there is no any option to speed up my computations? Therefore I will not be able to finish all the computations before the end of my PhD next year ...

After more than two weeks of 16-node calculations, I could check only 19 out of my more than 300 structures.

[antti_karttunen]

Hi,

I think the systems are not that huge and they should run rather fast if you are using RI-BP86. Only 19 displacements in two weeks sounds really slow. Some issues:

1) E5-2687W CPU has only 8 physical cores and 16 cores if you are using Hyperthreading. Usually using Hyperthreading is a bad idea when doing real computation on all cores (you may see small speedup when going from 8 -> 16, but definitely not near 100% - actually, you may see a slowdown due to CPU cache issues). So, I recommend that you should only use 8 cores.

2) You don't have much RAM, so you could set $ricore to 0 (just to make sure that ridft is not swapping to hard disk. Modest $ricore does not lead into such huge speedups anyway. I'm assuming you are using RI-DFT!

Best,
Antti

[Glxblt76]

Dear antti_karttunen,

Thanks for your suggestions. But the point is that I did not use RI. I used BP86-TZVP without the resolution of identity because we compared some of our structures modeled with and without RI and found out that the RI approximation introduced significant differences between the structures.

So, if I use only 8 cores instead of 16 it should speed up my calculations even without RI. Do you have any other suggestions in my case?

I bring some detail: I completed the frequency calculations for 19 molecules, each of which corresponding to around 200-400 displacements.

[uwe]

Hello,

I used BP86-TZVP without the resolution of identity because we compared some of our structures modeled with and without RI and found out that the RI approximation introduced significant differences between the structures.

this seems to be an interesting case - could you please send one of the coordinates where this happened to the Turbomole support, please? It is very important to check if that is a technical problem or a problem with the RI approximation.

Just send the coord file to turbomole@cosmologic.de

Thanks,

Uwe

[antti_karttunen]

Ah, OK, I misread your original post. 19 molecules out of 300 is definitely better than 19 displacements out of 300.

I definitely recommend at least testing the calculations by using only 8 CPUs instead of 16. Then you will see, what kind of effect Hyperthreading has.

Finally, I think RI-DFT could make even an order-of-magnitude difference here. Please contact Uwe about the problematic cases, I've never seen a case where RI-DFT vs. DFT results in large structural differences (as far as proper auxiliary basis sets are used).

Best,
Antti

[Glxblt76]

Just send the coord file to turbomole@cosmologic.de

Dear Uwe,

Unfortunately, I'm under confidentiality agreement.

What I can tell is that the difference between RI and no RI was not large and we did not notice it for all structures. For example, for some surfactants, the alkyl chain rotated to 30° but we noticed no difference in hydrogen bonds or covalent bond lengths. To be brief, only very flexible dihedral angles could rotate significantly between RI and no-RI structures. We discarded RI because since it is an approximation, it seemed conservative to us not to use it, but we did not considered it erroneous.

@antti_karttunen: thanks for your kind answers, I will follow your recommendations for my next frequency calculations and see if there is some difference then

[uwe]

Hi,

you do not have to post the coordinates here in public. One of the nice things of Turbomole is that support is not done via a forum or mailing list. It is usually not a problem to share confidential data once we have a non-disclosure agreement. Please ask your whoever-is-responsible-for-it to get in contact with the support team.

Using RI will in your case speed up the calculation by a factor of 10 to 30. So instead of having done 19 out of 300 calculations, all jobs would have finished already... I think it's worth looking at this RI phenomenon since it could be a hint that something else did go wrong. And you wrote that it's time critical. But it's of course all your decision.

BP-TZVP and COSMO sounds like an input for COSMOtherm. Is that correct? In that case I'd recommend to compare the resulting sigma profile and potential when using RI and non-RI cosmo files of those cases where the structure is different. If they are identical (that's what I assume), you will have exactly the same accuracy and get the same properties within COSMOtherm - no matter if you have used RI or not.

Regards,

Uwe

[shafikoff]

Hi

I was trying to use NumForce to calculate the excited state vibrations. However, when using RI-M062X/def2-TZVP, dscf end abnormally because of $rij flag. Calculation without the $rij proceeds quite fast until the 64th step out of 344 of gradient at displaced geometry calculation then gets stuck. How could this be resolved?

Marsel

[uwe]

Hello,

did NumForce stop completely? Just restart it, it checks for already finished steps and just runs the missing ones.

If you have an RI-DFT input with $rij in the control file, use the option -ri for NumForce to let it run ridft and not dscf.

Off-topic: Recently I had problems with stability of M06-2X already for ground state densities (running TDDFT afterwards), and this functional is known to be a bit tricky in some cases. If just some of the points are not converging or show strange results, I'd give a more stable (and for TDDFT on average also more accurate) functional like PBE0 a try... But this is just my 5 ct....