Parallel execution of ADC(3)

marand71 · September 8, 2021, 11:23am

Dear users and developers!

I would like to ask you how to make qchem (I’m using the 5.1 version) compute ADC(3) excitation energies in the parallel mode.

I understand, that it has to be done within a single node with a number of cpus (cores). The program is compiled with openMP, I am using the recommended switches (-slurm -nt … ) and it runs fine while doing the SCF, yet as soon as it goes to the correlated part (MPs, ADC iterations) it immediately switches to a single-core job.

Please, assist me here as I have not a clue what I should do. If you need information about the environment, let me know and I will try to fill you in as best as I can.

My very best regards
Marcn

epif · September 9, 2021, 5:42pm

Sounds like you are already launching Q-Chem with multiple cores via qchem -nt X and the problem is that you do not observe any speedup with ADC calculations. There are several recommendations:

Verify that there is in fact no speedup by running with a different number of cores (for example, -nt 4 and then -nt 8 provided that your compute node has 8 CPU cores or more). How do the overall timing and ADC timing change?
If SCF seems to speed up, but not ADC, it could be caused by its being limited by disk I/O bandwidth. Make sure you are providing sufficient MEM_TOTAL. There are some suggestions for CC jobs 6.16 Memory Options and Parallelization of Coupled-Cluster Calculations‣ Chapter 6 Wave Function-Based Correlation Methods ‣ Q-Chem 5.4 User’s Manual that are also applicable to ADC jobs.
If the job is large (say, over 500 basis functions on a 128GB node), using RI will reduce memory pressure and speed up the job.

marand71 · September 13, 2021, 4:28pm

Yes! It does work fine now, when I migrated to a node with 760GB of memory. Thank you very much for the helpful advice. I will have another question concerning the performance of ADC(3) for the totally symmetric excitation and for the triples states, but I will post in with a separate topic, as it might be interesting for others as well.
My very best regards!
Marcin Andrzejak