Reducing Walltime

[SAcompuser]

Hello all!

So I'm running a dft dynamics run of a glucose molecule and I'm noticing a significant difference between CPU time and Wall time, with the former being about 7.7 hours and the later being 11.8 hours when I run NWChem on 32 cpus with MPI. Is there any way to decrease the overhead to decrease the difference? I have attached my input script below.

Any advice would be greatly appreciated.  

[Edoardo Aprà]

One  cause of slow-down is I/O.

The following changes could reduce time spent in I/O operations:

1. I would switch the SCF from semi-direct to direct by changing the line

semidirect filesize 100000000

to

direct

2. The trajectory file is written by default at every timestep. By changing the value of print_xyz the trajectory file will be written less often.

For example, the following input line will cause the trajectory file being printed every 5 time-steps

qmd

 print_xyz 5

 ...

end

Your memory line seems too large to me. Do you have 3840 (120*32) gigabytes of memory on your system?

(Do you see any swap memory used while your job is running?)

I would change the line to more reasonable values, for example:

memory  stack 1000 mb heap 100 mb global 500 mb noverify

One more tip: switching from the default cartesian to spherical basis functions should speed the calculation, too.

https://nwchemgit.github.io/Basis.html#spherical-or-cartesian

basis "ao basis" spherical

[SAcompuser]

Hi Edoardo,

Thanks so much for you advice.

In terms of the memory, I have 120 gigabytes in total on that computing node. I thought I was assigning the total memory for all MPI process associated with the "mpirun -n 32 nwchem glucose.nw" command, it that not the case?

[Edoardo Aprà]

The memory input line specifies the amount of memory PER PROCESSOR CORE that NWChem can use for the job.

In your case you do have 3.75GB=120GB/32 per core.

I would not exceed 3.00 GB in your request (80% of 3.75GB)

https://nwchemgit.github.io/Memory.html

[SAcompuser]

Okay I will correct that. 

Thank you very much for your advice Edoardo.