Hi guys,
I am trying to optimize two semiconducting CNT molecules by using Gaussian 16.
My goal is to calculate interaction energy of the two CNTs.
The total number of atoms in the system is ~ 360.
Attached is the Gaussian input file.
(The molecular structure from the file is distorted because the optimization is ongoing.)
However, it has been about 2 months using 68 cores but it is not optimized yet.
I thought that it would take forever, so needs an approximation.
I freeze one of the two CNTs and the other remains free, and am still calculating it (refer to the attached file).
Is there any other efficient way to optimize this system?
Please help me.
First, check the initial structure (the system charge is neutral and check the bonding capacity of atoms, carbon has four capacities to form bonds). I checked your structure with Avogadro and as you see, when I used the option to add hydrogen, hydrogen was added to your structure. This means that the capacity of the carbon to form bonds was not completed in some places and some carbons had three bonds. So your initial structure has a problem and for that, the optimization is prolonged and your structure is distorted during the optimization.
Please check the structure again.
Second, it's better to do an initial optimization by the Molecular Mechanics force field (by Avogadro) and then start optimization by Gaussian. Instead of using an approximation, I suggest doing optimization with a lower method (Semi-empirical: PM3, AM1,...) and after optimization, prepare an input file from that output file, and do optimization with the DFT method.
Check the attachment.
For such large systems, usually plane waves work better than the gaussian basis set. Pseudo-potential programs such as Quantum Espresso or Vasp can do both tasks, structural optimization and interaction energy calculations, more efficiently.
Dear Chaneui Park,
From my experience, I suggest you to optimize with molecular dynamics first (Avogadro for exemple) to have a good guess your calculation. Then using Gaussian you can perform a PM3 calculation first and then after you increase more and more you level of theory until reaching of your goal. With the use of this method, be careful to check frequencies after each change of method.
First, check the initial structure (the system charge is neutral and check the bonding capacity of atoms, carbon has four capacities to form bonds). I checked your structure with Avogadro and as you see, when I used the option to add hydrogen, hydrogen was added to your structure. This means that the capacity of the carbon to form bonds was not completed in some places and some carbons had three bonds. So your initial structure has a problem and for that, the optimization is prolonged and your structure is distorted during the optimization.
Please check the structure again.
Second, it's better to do an initial optimization by the Molecular Mechanics force field (by Avogadro) and then start optimization by Gaussian. Instead of using an approximation, I suggest doing optimization with a lower method (Semi-empirical: PM3, AM1,...) and after optimization, prepare an input file from that output file, and do optimization with the DFT method.
Check the attachment.
The performance of the optimization method in Gaussian16 depends on several things. The first one is the method itself: if you want a quick method, then probably HF would be a solution with the smallest basis set. As somebody mentioned here, it would be good to preoptimize your molecule using Avogadro. In Avogadro you can select the force field for this purpose, for example, GAFF or UFF. Another thing which affects the performance is your hardware. If your hardware is not good enough then things will go slow. Gaussian16 has GPU parallelization which can speed up the calculation. If you have a good machine with the big enough RAM then you can speed-up your calculation by giving the right amount of memory, by specifying it in your Gaussian input-file. Then the last thing is interconnect: if you have fast enough interconnect in your machine you may get your calculation running faster.
If you are using super-computers with other users then it is worth to check if you are the only one who is using the computational node. It is better to reserve whole nodes even if you are not going to use all cores from the nodes. Also on super-computers you need to check if your big temporally files are stored on computational nodes and they do not accumulate on your storage on cluster. Such a setting can also improve performance of your calculation. This setting you do in your job-submission file.
Inna Ermilova The GPGPU speedups advertised by Gaussian(!) are very modest, a factor of 2-4. https://gaussian.com/g16/gpus.pdf
Since this is an advertisement, the actual speedups may be even lower. Also, GPGPUs tend to be very expensive, a single card costing thousands of dollars. But, GPGPU setups often don't have one card, but several.
In fact, Gaussian's ad states "The timings compare running across 28 CPU cores on a DGX-1 server (Broadwell CPUs) versus running on the same server across 28 CPU cores and 8 GPUs (with 8 CPU cores serving as GPU controllers during the GPU-enabled parts of the calculation)"
Their benchmarks are with eight(!) Tesla V100 cards with 16 GB of memory. According to Amazon, https://www.amazon.com/PNY-TCSV100MPCIE-PB-Nvidia-Tesla-v100/dp/B076P84525, these are not cheap cards: with one card costing close to $8k, the GPUs cost over $60k in total.
For $60k you can buy a whole lot of CPU cores; either in a single machine, or even set up a computing cluster. The speedup over the original 28-core setup would be way bigger with CPUs for the same amount of money.
Susi Lehtola, this is not an advertisement, this is the way the software can run on parallel. It has both CPU and GPU parallelization. I have tried G16 on GPUs from Nvidia Quadro cards up to Volta100 and I can tell you that the performance is fantastic. You can not compare them with CPUs. The speedup which you mention strongly depends on the molecule(s) which you are going to calculate and selected methods. All molecules do not run equally fast/slow. Then since the question was asked about how to run things faster it is important to mention GPUs. Yes, you can buy whatever you want with your money, but we, scientists, are using the dedicated computational centers which are supported by governments and other funds. Such centers might have various CPUs as well as GPUs. We do not pay anything extra for hardware or software (we do not order anything from Amazon): everything is prepaid. Then I do not know what kind of options the Mr Park has at his place, so all what I can do is to give him several options and he can decide which ones can be suitable.
Inna Ermilova a speedup of 2-4x with a cost of 200-400x is not fantastic, and the advertisement's speedups are optimistic scenarios. Compute centers often operate on a credit system, and there the same argument holds: GPU nodes are much more expensive to run on than CPU nodes, and if the speedup you get from a GPU is not reasonable (you should see speedups at least of the order of 10x -- 100x, not 2x), then you should not bother trying to use a GPU.
Susi Lehtola, I had much higher speedup than what you say here. I really do not understand what goal do you have with your arguments here? Why should we care about how computational centers operate and about their economy? We apply for projects at different centers with different resources, we pay nothing. They give us free hardware, free software and even service for beginners, i.e. they can give you test-scripts if you aren't able to figure out what settings are the most optimal. So what is your problem exactly here? I will try whatever I have access to and whateer I want without asking your permission or advice. ;-) So every other scientist should do: as a scientist you must try if you want to be successful with your research.
Inna Ermilova What's your problem? I am looking at the official numbers by Gaussian's advertisements (link above), and looking at the cost efficiency. You claim "you have much higher speedup" but don't even show any figures.
It is every computational scientists' responsibility to ensure they are spending their resources wisely. If you have completely free use of compute hardware, good for you. However, most people have to queue and ask for computational resources. GPUs are much more expensive than CPUs, and this is reflected in the typical credit systems. It's better to use GPUs where they are actually beneficial, instead of silly toy applications where the costs far outweigh the benefits.
Susi Lehtola, they have a help center, you can get in touch with them if you have questions about their software. I was running calculations before on various graphic and I was very satisfied. I do not show my figures here, but I submit them when I write my proposals for CPU-time. I will never give figures to a random person (even if you are an established researcher) who will later be my competitor for the resource. You want figures, you set up tests and then you make your own figures! Otherwise, the one who has figures wins CPU/GPU-time. :-)
Do not worry, we are spending our resources wisely. I guess the biggest concern here is the scientists who do not know how to set up the efficient calculation and who complain that things do not work. One can be a good quantum chemist but can be bad at programming of parallel computers and efficient job-submission. GPUs are more expensive, but I repeat it for you: we have them at computational centers and we do not pay anything for them, neither for the software. Governments and scientific councils buy things for our research and give it to us for free based on if we have research grants.
Inna Ermilova you are making empty claims unsupported by data. Maybe you are getting paid by Gaussian to advertise their software, but you should not mislead other people into spending huge resources for nothing. I again restate that this is an official ad by Gaussian that shows their scaling: https://gaussian.com/g16/gpus.pdf
which is downright horrible for the cost of a GPU.
Dear Susi, my claims are not empty, I got computational time for them in several centers in Sweden where I have provided the benchmarking. Benchmarking is required in most of computational centers, but not so many people can do it properly. No I am not getting paid by Gaussian, but if you read the actual question, you would know that it was about how to calculated faster with Gaussian-16. All what I have done is that I answered the question about the matter. Writing irrelevant stuff would be rather irrelevant. Moreover, your way of talking on Researchgate with accusations towards me (i.e. saying that I am working for Gaussian) is impolite and silly, taking into account that we both have Internet access and find out easily who is working for whom. I think you shall work on your reading and manners in general. (This is not Facebook, dear.)
Also if you read the paper which you are referring here too, then you would also see that the paper is about Alanine 25 and GFP (which are very large molecules) where they use the particular settings (see headers of files in column 2). I was not running for these two species, I had other large molecules of the interest. For my molecules I have obtained the desired result. Making the software and taking the money for something would not work would be stupid from Gaussian. As a user I was very satisfied, particularly, after my experience with Gaussian-03 and Gaussian-09.
I do not mislead anybody, this is you who did not read the report which you are refering here to and you didn't manage to read the question either which was about Gaussian-16. Instead you blame me... I would strongly suggest you to take a rest and to take a good lunch, perhaps, things will get better back there. :-)
Inna Ermilova Lol. Documented Gaussian 16 speedups are here: https://www.researchgate.net/deref/https%3A%2F%2Fgaussian.com%2Fg16%2Fgpus.pdf
This is advertising, it is in the company's interest to make the software look as good as possible, meaning the results are probably more optimistic than they are in practice; and this is still Gaussian 16. You have still to provide any evidence to your claims. Anyone can claim whatever they want, but in science it is data which has the final say and you have presented none. Since this discussion is obviously futile, I am not wasting any more of my time on it.
Susi Lehtola, for the simplicity's sake, can you just read the actual question which Mr Park have written? Then what software does he mention? Name it, please! :-)
Your document includes results only for 2 systems and for 4 set-ups. As an experienced computational researcher you should know that quantum chemical methods are way more broad than this document. They showed a small example which can not be used as a result for all systems and methods.
I will remind you again: results of my performance tests are provided to computational centers and they granted me CPU/GPU-time on various computer including the latest graphic cards.
I agree with you that the discussion is getting more and more ridiculous, it wouldn't go this way if you read the question and had a sense of behavior on a public research resource. :-)
Inna Ermilova I've said Gaussian 5 times, and Gaussian 16 another two times. Now, 7 and 3 times. Which software did I mention? Hint: it was Gaussian 16.
You're not serving your cause by refusing to back up your claims that I am wrong. You have not refuted the data that is available and that I have presented, instead you have claimed your own experiences, but have not quantified them in any way. Is this how science works in your field?
The document is not mine, but official scalings reported by the sellers of the software. One of their example was Alanine 25, but the other was Green Fluorescent Protein that has 238 amino acid residues and ~4k atoms. The OP's structure only has 360 atoms, which is not very big, meaning that the scaling one will get from GPUs will not be noticeable.
It's highly unlikely that someone has empty nodes with >$60k worth of GPUs; by investing in more nodes, instead, you can increase the actual performance easily by 10x-20x. And if a computing center already has GPUs, they typically "bill" you in relation to how much the hardware actually costs: CPU nodes are much cheaper than GPU nodes, meaning that if you just double or quadruple the number of CPUs you might get the results faster than with GPUs but at much smaller cost. Using GPUs is not a panacea, you should benchmark the performance first to see whether the speedup is worth the expense. Same goes for CPUs: you need to check how far the program scales for your problem.
Although publishing Gaussian benchmarks is forbidden by the license conditions, it is not generally known to be a fast, well-scalable code. There are many free packages that *do* scale up to large calculations, e.g. NWChem. While this may not be the solution to the OP's question, it is usually good to think out-of-the-box once in a while; note that some of the other answers here have also included using other software for specific tasks.
Chaneui Park it looks like your initial geometry is very far from the minimum.
In addition to using force field minimization in e.g. Avogadro to get a better initial geometry, modern tight-binding methods are also extremely useful for this purpose in contrast to older semiempirical methods like PM6 which are trained for the estimation of energies but not geometries.
For instance, Grimme's xtb program is blazing fast and can handle systems of very large size
https://www.chemie.uni-bonn.de/pctc/mulliken-center/grimme/software/xtb
After a force field minimization, you might try using xtb to get the geometry closer to the DFT minimum.
Chaneui Park note also that B3LYP is probably not the right method for this task: dispersion effects are likely to have a large effect on the energy and geometry of the CNT system, so you should consider using a dispersion corrected functional like B3LYP-D3.
Susi Lehtola, so finally, you understood that the question was about Gaussian? Or do you still think that it was me who suggested the software?
Regarding my cause and you: my cause is served where it has to be served and I get the computational time for that. Your private opinion does not matter for me. The reason why I wrote the answer is because Mr Park has asked the question, not to please his majesty Susi. ;-) He will read the answer and decide what to do out of his resources. No more than that. He doesn't have to buy anything.
If you were an expert in QC-calculations you certainly would not rely on two examples for few methods. Those speedups are provided for certain cases which they mentione in the table. They didn't generalize in their report that the result will be the same for all methods and molecules, because they know that it will not. They have provided two large molecules in order to show that there is a scaling for such systems as well.
Regarding the usage of CPU/GPU-time: I do not know how it is counted at the place where you calculate your stuff (if you calculate at all), but in the places where I use to run the scheduler simply counts only CPU-hours, not GPU. So if a user will reserve one graphic card with one CPU, he will get the results as fast as if he would reserve more than one CPUs. This simply means that if you use too many cores instead of one GPU you will run out your monthly quota for CPU-time, what will not happen to a user using GPUs. Since I do not run my calculations systems mentioned by your pdf-document (they are a bit smaller than that and methods are different), I win a lot using GPUs with Gaussian-16.
Again: discussing the scalability you shall talk about particular methods and systems. You can not generalize the opinion about Gaussian-16 based on two examples. This is very wrong. So if you write your research like that I am not surprised that you are so nervous and stressed even on public resources like you are here. (Just think that all your impolite and harassing behavior towards me will be read by other specialists! I didn't want to be the one to remind you, but this is how it is. It is very sad to see the behavior like yours, considering that you work in academia.)
There are many packages indeed out there, but Mr Park has mentioned Gaussian-16 in his question. One can write a random crap regarding that here. Moreover, Gaussian-16 is known to have more options for methods, basis sets, solvent models etc. Since the price for the whole campus or a computational center is rather low, a lot of people use it. Also one can make input files using other free programs without investing in GaussView.
Inna Ermilova I still don't understand where you got the idea that I was talking about anything else than Gaussian, since IT IS THE ONLY PIECE OF SOFTWARE I WAS TALKING ABOUT, and I REFERRED TO GAUSSIAN-16 DOCUMENTATION.
You're fighting windmills here.
Susi Lehtola, from what you say now and in your previous comment I see that you have no idea about what you are writing here. Firstly, you were blaming me for talking about Gaussian16 and then you were accusing me that I am trying to sell it here, because "they paid me", according to you. Now you are writing other weird things. I can only conclude that considering the situation with covid-19 around the world it is normal that scientists like you got affected mentally. However, it would be better for you to visit a psychologist (if it is opened in your country or make a zoom-meeting) in order to discuss problems with your behavior.
I do not fight you, this is you who is writing harassing comments to me. My comment was written to Mr Park, but it made you boiled from toe to top. I have to repeat this for you again: this is not Facebook. This site is dedicated to the scientific community, so you shall behave if you do not want to be embarrased meeting everyone live when corona is over. ;-)
Mr Park, I hope that you are fine.
I am totally in agreement with previous answers. Try to organize better your input file and after, re-optimize in a lower level. After you have a minimal with a lower level of theory, try to perform the calculation with a better level of theory and basis set according to the literature.
I have some cores that could help you if you want.
Kind regards, Julián.
Inna Ermilova and Susi Lehtola
Please, keep calm! Don't fight!
For this, exist internal message.
Most of the discussion is fruitful but the question of Mr park is specific!
Remember that this medium is for sharing knowledge and for scientific discussions, no for scientific egocentrism!
A very big hug :)
Julián.
Julián, I completely agree with you and this is what I was trying to explain to the guy. No fights from my side. :-)
- Badges
- Science topic