I've been fooling around with the various ways that we're running NWChem on our local HPC and one particular thing became apparent that I'm not sure how to address.
I'm doing DFT calculations on a number of different molecules.
When I ran one particular job with approximately 25 other molecules also using NWChem (submitted separately through separate scheduler jobs on separate nodes) it took us 8 times longer (wall time) to run the job than when it ran alone. The cpu time was 22 times longer when run with 25 other jobs than when it ran alone.
cputime (run with 25 other instances):128931.7 (run solo):5801.8
wall time (run with 25 other nwchem): 166354.1 (run solo): 20022.9
I thought that NWChem is supposed to be running separate instances, but it seems like there must be some interactions between the different instances of NWChem. This is a HUGE difference in run time for the exact same file. (but at least the results came out the same!).
I'm sure there is something that I'm missing, but don't know what it is!
I have copied an example of our NWChem input below:
echo
memory 2 gb
scratch_dir /lustre/groups/millergrp/scratch03
permanent_dir ./the_good_stuff
start 2ov3
title "dimermer of ovalene trimer"
charge 0
geometry units angstroms print xyz noautoz
symmetry C1
C -2.85591975 3.66567123 0.00000000
C -3.54223808 2.48587843 0.00000000
C -2.85839089 1.22452898 0.00000000
[cut lots of atoms here...]
H -2.24426423 3.40783451 -13.52506403
H -3.97569027 4.38673402 -12.06476140
END
BASIS
* library 6-31G*
END
DFT
XC B3LYP
MAXITER 100
print "final evals" "final vectors"
END
TDDFT
CIS
NOTRIPLET
NROOTS 50
END
TASK dft
|