Restarting Jobs for LBFGS optimizers

Vasp transition state theory tools

Moderator: moderators

Post Reply
sidd20111992
Posts: 8
Joined: Tue May 16, 2017 4:09 pm

Restarting Jobs for LBFGS optimizers

Post by sidd20111992 » Mon Jul 17, 2017 1:42 pm

Hi vtst community,

Our group makes use of a queue which has a short queue time of 4 hours. Because of this sometimes our jobs don't do sufficient number of geometric steps to converge in just one run. We were wondering if this could somehow affect the efficiency of the LBFGS algorithm, which requires certain number of geometric steps to built the hessian. If yes, is there a way to read the steps from previous runs when we restart the run, such that the algorithm now considers the steps from the run before the restart to built the Hessian as well and not redo the process every time. Awaiting your reply.

Thank you

Siddharth

graeme
Site Admin
Posts: 1622
Joined: Tue Apr 26, 2005 4:25 am
Location: University of Texas at Austin
Contact:

Re: Restarting Jobs for LBFGS optimizers

Post by graeme » Tue Jul 18, 2017 12:24 pm

We do not, but I think that it's a good idea. The optimizer could write displacement / force pairs in a history file which would get read in upon restart.

sidd20111992
Posts: 8
Joined: Tue May 16, 2017 4:09 pm

Re: Restarting Jobs for LBFGS optimizers

Post by sidd20111992 » Tue Jul 18, 2017 1:24 pm

Hi Professor Henkelman,

Thank you for the quick response. Can you please give us some tips in form of areas in code that we could try to modify so as to create this history type of file and then further read it in upon restart. Awaiting your reply.

Siddharth

Post Reply