December 2011:
Tried running on NERSC Carver with mkmem 0 (an not) in scratch. Tried using bigmem 12 node processors with more memory per processor, but always seemed to run into memory problems either within core or in workspace.
2/15/2012:
First attempt to run in XSEDE on Blacklight.
Some key input variables:
mkmem 0
chkprim 0
ecut 35.0
ngkptp 6 6 6
nshiftk 4
nstep 150
toldfe 1e-7
The input.sub file was modeled after documentation here
ncpus=176 to accommodate the 171 k-points
walltime=48:00:00
output:
The run was terminated after 21 iterations because the wall time was exceeded. The energy had converged to -1463.4306 Ha (to 4 decimal places, so 3 decimal places from the toldfe condition). This also cost about 7000 SUs, which is a lot.
2/23/2012
The run above indicates that things are working with this set up. However, since the run cost so much time we decided to reduce the number of k points by no longer shift the k-point grid four times. This run had an nshiftk=1 and shiftk= 0.5 0.5 0.5, but everything else the same as before.
ncpus=32 to accommodate 27 k points. Walltime=96:00:00. Run on Blacklight.
output:
The run terminated after 33 iterations. According to the system output file, it did not exceed walltime but may have been kicked out by a debugging program on blacklight…we don't know exactly what caused it to terminate at this point.
3/14/2012
To test if the program is the problem, the input, or the computer, I re-ran the shiftk=1 input (above) on NERSC's Carver.
But I can't get the program to actually run due to memory issues!
6/1/2012
Finally get an output on blacklight XSEDE after 34 iterations. I chose to set the nstep to 34 because the job died after 36 iterations the last two times I ran it. Even though there was no walltime error, from monitoring the time as the job ran I believe it was the 96 hour walltime which forced the jobs to end. I think it is fine that the toldfe of 1*10^-7 was not quite reached in 34 iterations because the last three energies were quite close (The last two energy differences were 1*10^-8 and 1.1*10^-7):
Etot: -1463.4307067938
Etot: -1463.4307068038
Etot: -1463.4307066941
Some key input variables:
mkmem 0
chkprim 0
ecut 35.0
ngkptp 6 6 6
nshiftk 1
nstep 34
toldfe 1e-7
The input.sub file was modeled after documentation here
ncpus=32 to accommodate the 27 k-points
walltime=96:00:00
ELFs created using documentation here
Iterative Hirshfeld charges found following documentation here, and were found to converge after 86 iterations