Memory Usage During Optimisation

[Peter Gillespie]

Hello Everyone,

I have been running geometry optimisations of some quite large systems 
(~2000 electrons after pseudopotentials are used) with a hybrid HF/DFT 
functional (HSE06) and have been trying to fiddle with the HFx memory 
allocation to maximise performance.

Using a combination of ADMM and the psmp executable on Archer with a lot of 
the available memory allocated to HFx, I have managed to bring the time per 
SCF step down quite considerably. However, the optimisation then will not 
run beyond a few steps before the program runs out of memory. Leaving even 
2GB per MPI process for the rest of the program leads to it eventually 
running out of memory, and I have found that I cannot allocate more than 
half of the available memory if the geometry optimisation is to work for 
any reasonable length of time when using a CG optimiser.

The CP2K manual implies that the different geometry optimisers (BFGS, 
LBFGS, CG etc) are best suited to different sizes of systems, but does this 
mean that some are more robust than others for larger systems, or does it 
mean that some are more efficient (and potentially less memory intensive) 
than others? Also, is there a good way of estimating how much memory the 
geometry optimiser will likely use, or any factors that typically affect 
memory consumption besides the system size alone?

Best Regards

-Peter
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20170921/f673f8c1/attachment.htm>