Dear,
Actually I am performing a sp CCSD(T)/aug-cc-pvQz (in gaussian09) with around 720 basis functions and 1100 primitive gaussian of size (the compound has pseudo-potential core for a metal) on a TS structure. With the basis aug-cc-pvTz the time of calculation was very fast as was expected and with an other system very similar in size and system only delay 1 week with less cpu indeed.
The problem is that with 48 cpu 220GB RAM (required memory was 650GB) the calculation is taking a long time in the last step(when the system is solved and appears "Largest amplitude=" of the CCSD(T) calculation (now is in more than 1000 hours!).
Are there some consideration, suggestion or explication of this trouble? Is it possible to solve it?
Thank you very much,
Joaquim Rius
I am not really an expert on this, but you mention “650 GB required memory”. If your computer does not have enough RAM, it sends the integral tables to the hard disk drive, either to the swap partition or to some named files, and that really slows the machine down (and kills the HDD). So you should check the disk activity. If it is high, add RAM or, if this is not possible, use HDD in RAID 0 configuration.
Dear Ulrich,
Thank you for your answer. The problem is that in the cluster I think that I am not available to add more RAM (but I will ask to the technician).
It is possible that could be better use only hard-disk than a mix of RAM and hard-disk?
I do not know that strategy of Gaussian09 and therefore cannot say whether a mixed set of storage devices (RAM and HDD) accelerated or slows down the program. According to Murphy's Law, however, the most frequently needed set of data goes to the slowest device, and this therefore determines the overall speed.
Dear Joaquim,
did you try assigning read-write files for your job as described here: http://gaussian.com/g_tech/g_ur/k_link0.htm
Maybe this could be helpful, too?
Best regards,
Aleksey.
Dear Krzysztof,
Thanks for answer. A question that I have, it is reliable to perform the optimization with Gaussian for example, and then, run the sp-CCSD(T) with other? I say that because sometimes I try the same calculation with Gaussian and Gamess, and the energies was slightly different and in my case, I am performing a high accuracy calculation and I do not know if this difference will imply an increase of the error.
CFOUR program is also a good choice to CC calculations since it was developed to this unique purpose, it is free and the input is very easy to build. Gaussian, as said, is not well suited for this particular theory.
You can also consider using DLPNO-CCSD(T) in Orca. This is a near-linear scaling method developed in F. Neese's group, and it seems to be very efficient even for large molecules; besides, Orca apparently scales quite well with many CPUs.
Dear Amy,
This calculation is a single point energy. The problem is that similar size calculation stay only for 1 week, this one for more than a month stacking at the same step, the last one with the iterative resolution of the equations system.
Also, I have the calculation with aug-cc-pvTz and it was perfectly calculated.
Dear Joaquim,
In terms of single point calculation, I think MOLPRO is one of the fastest program for CCSD(T) calculations. I believe the absolute energy of the same geometry from Gaussian and MOLPRO are different, but I'm not sure if you care about the absolute energy. There are lots of cases where people actually only care about the relative energy, for example, the dissociation energy and the reaction barrier. In these cases, all you need is to perform two single-point calculations and calculate the energy difference. And for these purposes, the results from MOLRPO are as reliable as Gaussian.
So my suggestion is that you consider using MOLPRO, as it is much faster than Gaussian for CCSD(T) calculations, and as pointed by Krzysztof, F12 method is implemented in MOLPRO so that you can use a smaller basis set with CCSD(T)-F12 to achieve almost the same accuracy of CCSD(T) with larger basis set.
Best,
Chen
- Badges
- Science topic