I’m running a simulation in which there is a rigid body. I ran a simulation in which I keep the rigid body immobile by using force off and torques off in ‘fix rigid’.
After I equilibrate this system, I launch a new simulation which starts from the equilibrated system but I let the rigid body loose to wander around. The simulation runs for a few thousands timesteps as normal but it suddenly slows down such that each single iteration takes a very long time to progress it’s almost stopped! None of the thermo output that I get (temperature, pressure, center of mass position of the rigid body) show abnormal changes nor illegal values like NaN I see in the outputs.
I thought maybe the rigid body crossing the periodic boundary can cause this behaviour, but although the rigid body is close to the periodic boundary it’s not crossing it yet when this slow-down happens.
If I turn off the force and torque of the fix rigid again (basically keeping it immobile) the simulation goes on smoothly.
I’d really appreciate it if anyone can help me to solve the problem, or hint me how I can track this slowing down behaviour.
3rd paragraph correction:
my bad! my guess was right, it happens slowing down happens after the very first atom of the rigid body crosses the periodic boundary. I write out the position and image flags of the rigid body and all flags are zero (rigid body is completely in the original box). So I believe the following note of the ‘fix rigid’ has been satisfied:
IMPORTANT NOTE: To compute the initial center-of-mass position and other properties of each rigid body, the image flags for each atom in the body are used to “unwrap” the atom coordinates. Thus you must insure that these image flags are consistent so that the unwrapping creates a valid rigid body (one where the atoms are close together), particularly if the atoms in a single rigid body straddle a periodic boundary. This means the input data file or restart file must define the image flags for each atom consistently or that you have used the set command to specify them correctly. If a dimension is non-periodic then the image flag of each atom must be 0 in that dimension, else an error is generated.
but as soon as the rigid body starts to cross the periodic boundary the stated slow-down happens.
i don't think there is much that anybody can do about it unless you
produce a simple example that reproduces the effect. i don't think
that the feature you are referring to as anything to do about it.
reliable information can only be had through proper profiling during
execution.
i don't think there is much that anybody can do about it unless you
produce a simple example that reproduces the effect. i don't think
that the feature you are referring to as anything to do about it.
reliable information can only be had through proper profiling during
execution.
thanks for the input, but i cannot reproduce the slowdown with 4 mpi
tasks on my laptop with a hyperthreaded dual core CPU.
i added a fix addforce 0.0 0.0 0.1 to accelerate the transition of the
c60 through the periodic boundary and cannot see any appreciable
performance difference outside of the usual slowdown after the initial
few steps due to the CPU powermanagement first overclocking the CPU
and then reeling it into the designated thermal envelope as it warms
up from the turbo boost.
there must be some other contributing factor. perhaps, your machine
has overheating problems?