memory errors

[esp]

i ran some 2 probe GNR devices just fine when they were only 10nm long, but I tried to do 20nm long and starting getting the pseudopotentials error .. I believe it has been translated to "out of memory" ... i reduced kpoints to 51 in the z direction, and i reduced number of history steps to 12, still cannot get it run with 4 nodes, 2.5GB memory per core ... this was running in a cluster with intel MPI .. i tried also instead of specifying pmem=2500mb, i tried a total memory allocation of mem=80gb (20GB per node) but then i got some abort error after some time

the funny thing is, i have a lab computer (not MPI) that i can run using mpich2, it has 24GB memory, .. it has no issues except it ran all night and got nowhere ...

it seems about 10nm long is the limit of what i can do  ... again, 10nm long device runs ok, about 2-3 hours for the first calc, then 1.5hours per IV point on a transmission calc ... anyone have suggestions given this information?

[Nordland]

starting getting the pseudopotentials error .. I believe it has been translated to "out of memory" ...

huh?

[esp]

I am referring to this error that starts:
"PSEUDOPOTENTIALS_PATH=$EXEC_DIR/. .... "

It has been posted on Quantumwise many times, I thought it was generally understood to mean "out of memory" ..

Which part of the question was unclear?

[Nordland]

Okay

It is not a pseudo-potential error. It is just the command ATK is executed with on linux.

I will look at little at it.

[esp]

I see thank you .. I was using a AGNR of width N=25 (25 atoms along width - armchair config in z direction), and length of 20nm ..

[esp]

latest error, length 16nm

+------------------------------------------------------------------------------+
| Left electrode chemical potential  =   -5.755239 eV                          |
| Right electrode chemical potential =   -5.555239 eV                          |
+------------------------------------------------------------------------------+
| Checkpoint Handler                                                           |
| Filename : /tmp/checkpoint60121649.nc                                        |
| Interval : 0.5 h                                                             |
+------------------------------------------------------------------------------+
rank 0 in job 1  node0656_48140   caused collective abort of all ranks
  exit status of rank 0: killed by signal 9

What does it mean?

[esp]

Update, on the same cluster, i ran 16nm long device, N=10, Device Huckel Calc finished ok ... seems any larger (width or length) gives memory error .. the thing is though, how much memory do you need for something like N=25, L = 20nm ?  I have seen papers that simulated 90nm, I would love to see this as well, but unfortunately it seems it is not possible if i cannot simulate L=20nm on a supercomputer cluster without memory errors ... any insight would be helpful ..

[Anders Blom]

The error message(s) do mean "out of memory". Have you tried 12.8.b1 and using MemoryUsage() to see how much memory to expect the calculation will use? See http://quantumwise.com/documents/tutorials/ATK-12.8/MemoryUsage/

[esp]

No i did not realize ..i will try ..

by the way, it has ran 18 hours now over 5 nodes, has not finished even one transmission calc ...