Hi,
I have a user at SDSC (www.sdsc.edu) who is trying to run a vasp 5.4.4 job on one of our clusters and it seems to want to grab an infinite amount of memory. I ran his job on one of our large memory nodes (> 2TB) with a small number of processors and it still failed ( I monitored the memory use as it ran)
The attached tar file contains the input files, OUTCAR and the standard output file
Thanks,
Jerry
Dear Jerry,
Unfortunately your system is too large, so I will not be able to run it myself and reproduce the issue. However, could you start by using the vasp_std executable, instead of vasp_ncl? Your system does not have spin-orbit coupling, so I do not see a need to use the non-collinear version. This should already reduce the amount of memory needed.
Could you also try setting LSPECTRAL=.FALSE.? This should reduce the amount of memory needed.
Finally, if you have access to VASP 6.x you can use the low-scaling GW method described here.
Let me know if this helps with your case. Kind regards,
Pedro
Hi Jerry,
another point to consider is also the number of k-points along c/z, which is 5 in your users case. I see the point that for that kind of calculations you don't want to waste precision somewhere else, but c-axis is already ~7 A long, so your user might also go away with a cheaper set (which could be tested with DFT or at least cheaper than GW simulations).
Hth,
alex
Thanks very much! The user is not licensed for vasp 6 but he will try the other suggestions.
The user has a question about MAXMEM: he wants to know if that is working since there is no mention of it OUTCAR: (I did point out to him that it is in the xml file)
Jerry
The user just wanted to verify that MAXMEM is actually being used.
Thanks,
Jerry
Dear Jerry,
As written here, MAXMEM is indeed read during GW calculations, which is what your user is running.
Kind regards,
Pedro