run_style/respa using multiple processors

Dear Sirs,

I would like to ask if there is any known issue/remedy for apparent instability of multistep integrator simulations using run_style/respa. Please find attached 2 output files for 1 and 5 processor setting in PEO equilibration runs. The 1-processor run completes successfully, except for slight increase in density compared to a single step integrator run. However, the 5-processor with max. 1.6 fs step run crashes instantly (not observed for equivalent single step integrator).
Best regards,

Osvalds Verners

log.multistep (2.08 KB)

log.singlestep (198 KB)

Dear Sirs,

I would like to ask if there is any known issue/remedy for apparent
instability of multistep integrator simulations using run_style/respa.
Please find attached 2 output files for 1 and 5 processor setting in PEO
equilibration runs. The 1-processor run completes successfully, except for
slight increase in density compared to a single step integrator run.
However, the 5-processor with max. 1.6 fs step run crashes instantly (not
observed for equivalent single step integrator).

​there is not enough information for ​detailed recommendations except one:

i would expect that for running r-RESPA, you would need to reduce the
neighbor list delay, as it refers to the outermost r-RESPA level.
delay*time_step should be kept at about the same value.

​beyond that, i would recommend to convert your restart to a data file and
then make the same test with the latest LAMMPS patchlevel version again.
your summer 2015 version is missing several updates and bugfixes and
improvements to neighbor list processing and diagnostic output.​

​axel.​

Dear Sirs,

thank you for the recommendations. I should add, however, that today I could not replicate the issue using the same input (no errors reported). Therefore, partially a cluster related issue could be assumed.
Best regards,

Osvalds