From: Kevin Chan (kevin_at_skblnw.com)
Date: Tue May 25 2021 - 08:39:57 CDT
I am trying to speed up an FEP calculation using cudaSOAIntegrate with NAMD 3.0alpha9 single-node path, however, failed by getting either
FATAL ERROR: Periodic cell has become too small for original patch grid!
Possible solutions are to restart from a recent checkpoint,
increase margin, or disable useFlexibleCell for liquid simulation.
FATAL ERROR: Low global CUDA exclusion count! (116807 vs 116813) System unstable or pairlistdist or cutoff too small.
The system has been tested with NAMD2.14 and NAMD3.0alpha9 without cudaSOAIntegrate and worked perfectly fine.
So I have tried to tackle the errors by
1. pairlistdist 13.5 to 16
2. margin 4, 8, 16, 20
3. LangevinPistonDecay 100 to 1000
None of the combination worked. Sometimes first 1 or 2 windows (lambda 0 to 0.001) got finished (500k+ steps), but I never got the middle windows (lambda > 0.1) running. I also tried the alpha8 version and had the same results. I set stepsPerCycle to 400 and pairlistsPerCycle to 40 all the time. The run command is simply "namd3 +p1 +devices 0".
I understand that FEP with CUDA speedup is still in alpha, so any experience shared will be greatly appreciated.
This archive was generated by hypermail 2.1.6 : Fri Dec 31 2021 - 23:17:11 CST