Page 1 of 1

Trouble with the 64-bit version of Dalton

Posted: 12 Sep 2016, 13:00
by marand71
Dear all!
I have been trying to do some CC2 calculations concerning the electric transition dipole moments to the excited states and the magnetic transition dipole moments between the excited states.
I have had some troubles doing that, but eventually I know how to write the proper input :)

Unfortunately, for the system I am currently working on even in the modest basis set (cc-pVDZ) there is a memory bottleneck, which cannot be overcome unless I use the 64-bit build of Dalton.

However, I have come across a peculiar behavior of the code. In the case of 32-bit compilation, it works just fine, crushing only at the response section for want of more memory.

If I use the 64-bit compilation (--int64) with the Intel compiler and MKL libraries, the CC2 (ground state) calculations diverge and already the initial MP2 energy is way higher than the proper value obtained with the 32-bit version of Dalton. This situation occurs for both the serial and parallel compilations. Curiously, it does not occur for all molecules. The problematic one was the N,N-dimethyl derivative of the pyromellytic diimide (C2v symmetry). For the pristine dianhydride (D2h symmetry group) this problem did not occur and calculations went as smoothly as with the 32-bit Dalton, only they did not crush at the CC2 response section this time for there was enough memory available with the 64-bit compilation.

I am quite at a loss at the moment as to what I may do to circumvent or overcome the problem. Any advice that would get me out of trouble would be most welcome.

Below details concerning the Dalton compilations can be found.
./setup --fc=ifort --int64 --cc=icc --cxx=icpc --mkl=sequential --prefix=/ss/dalton2016/ build64intel_mxdist2500_mkl-seq
Note that the parameter MXDIST was set to 2500, as the default value 0f 250 was too small and the calculations crashed having started the CC routines.

Yours sincerely!

Re: Trouble with the 64-bit version of Dalton

Posted: 12 Sep 2016, 13:36
by taylor
The information about your build is perhaps less helpful at this point than the output from your calculation that fails? Going to 64-bit integers implies a memory request of more than 16GB, which implies something approaching 1 billion CC amplitudes, or at least well over 500 million. That is a pretty big calculation...

Best regards

Re: Trouble with the 64-bit version of Dalton

Posted: 12 Sep 2016, 17:34
by bast
dear Marcin,
the probably best thing you can do at this moment is to try to create a test case which demonstrates the problem. when doing this, please reduce the system to the minimum possible (while still demonstrating the problem). this not only speeds up and simplifies debugging, it often isolates - "crystallizes" - the problem. this requires some work on your side but only then we can really do something about it.
best regards,

Re: Trouble with the 64-bit version of Dalton

Posted: 13 Sep 2016, 11:58
by marand71
First, I am going to post the output of the calculations that went wrong (divergence at the ground state CC2 step - pmdi_ch3_dz.out) and the calculations that went through this step fine and only crashed much later, owing to insufficient memory (pmdi_ch3_dz_ok.out). Hope that some reason for this CC2 instability may be deciphered from them...

As for more systematic tests with smaller systems, I am currently trying to perform similar calculations for a series of single ring species. If anything similar occurs, I will definitely write about it.

My very best regards!

Re: Trouble with the 64-bit version of Dalton

Posted: 16 Sep 2016, 13:10
by marand71
I have managed to test the behavior of the 60-bit Dalton for:

1. the N,N-dimethyl-PMDI with the C2h symmetry (Previously the symmetry was C2v)
2. Pyridine
3. Azapyridines

For 2 and 3 the program worked fine, for 1 it returned the strange MP2 energy value, and the the CC2 calculations obviously diverged.

Could it be the molecular size that matters here? For the non-substituted PMDI the 64-bit Dalton worked OK.