Hi, I am running some transition state calculations in Gaussian on a moderate system (~50 atoms including 1 d-shell metal). For this, my method includes:
#p opt=(calcall,ts,noeigen) wb97xd/def2svp
I am trying to speed things up on shared memory paralel execution cluster by using up to 64 CPU (%nprocshared=64). The thing is, I put %mem=450GB, which based on gaussian manual for large SCF, that needs even 3N^2 words, is sufficitent for ~45 CPU. Indeed, the PBS output shows, that the calculation used only mem=76 GB, but vmem=480 GB. Also in the job logfile, the number of CPU is reduced from 64 to 41 due to ecpmxn. I see, that this is due to insufficient vmem on the system that I allocated.
My question is, is there a way, how to lower the allocated memory that gaussian asks for, if it doesn't use it? This is getting very limiting for me as this restricts the use of many CPU for larger systems.
Thanks for any hints and helps.
1) You don't need 450GB RAM for DFT calculations, I'm 99% sure you can remove %mem directive and the default memory limit will work
2) DFT in Gaussian doesn't scale that well to run it on 64 cores
Dear Adam,
I have run the calculations for 90 atoms, including Iridium, successfully by giving 4800MB memory (which is good enough).
Please find below the initial lines of the input file for your reference.
%mem=4800MB
%nprocshared=16
%chk=qst2_PPh3_Trans.chk
# opt=(calcfc,noeigentest,qst2,maxcycles=700) freq=noraman b3lyp/genecp pseudo=read scf=(MaxCycle=700,xqc)
Reactant
0 1
%nprocshared is an obsolete directive, use %cpu=0-47 for 48 cores, 0-15 for 16 cores. If it’s Gaussian 16, of course.
scf=xqc is usually not necessary and takes a lot of computer time.
Thank you both, I realise my mistake. I know about the scaling of gaussian, although I heard both good and bad experience. I'm currently doing small testing to see how it scales in different routines I'm commonly using. If you're interested in the results, let me know.
For the memory question, I see that the used memory is the important value, while the allocated memory doesn't change much (maybe anything at all?). the routine, in which ecpmxn reduces number of processors is a minor process and thus doesn't affect the overall Elapsed time.
About scf=xqc I know it takes longer time, but in some cases it's needed for convergence. I use it only if first calculation fails.
And finally, I know about %cpu, but I'm not sure how would it go on the cluster with multiple users. %nprocshared is still valid and avoids problems in the queue system.
On a cluster the cores (threads) are allocated to the job by Slurm/PBS manager, Gaussian will take what's given by the manager.
- Badges
- Science topic