[AMBER] Accelerated MD -- cudaMemcpy error

From: Dmitry Suplatov <genesup.gmail.com>
Date: Fri, 12 Aug 2016 23:53:07 +0300


I want to run aMD of a fairly large protein.

The protein is a homo-4-mere, 1356 amino acids in total. The system -
protein in TIP4PEW with ions according to 0.1M - contains 163 705 atoms.

I implement the protocol according to the aMD tutorial (
1. Energy minimization in water with restraints on the protein (5000 steps
due to a large system)
2. Water shuffling with restraints on the protein (10 000 steps as in the
original protocol)
3. Energy minimization with no restraints (10 000 steps due to a large
4. Heating in NVT from 0 to 300K in 250 000 steps with restraints on the
protein (i`ve checked - no bubbles)
5. Equlibration in NPT - restraints are gradually removed in 12 steps
(-0.25 kkal/molA2 per step), 85 ps per each step
6. Equilibration in NPT with no restraints - 10 ns (density converges)
7. Free dynamics in NVT for 1 ns
8. aMD

The parameter file for aMD based on the freeMD is as follows:
Step-8: Accelerated MD in the NVT ensemble
imin=0, ! Flag to run minimization - No minimization
irest=1, ! Flag to restart a simulation - Restart the simulation,
reading coordinates and velocities from a previously saved restart file
ntx=5, ! Option to read the initial coordinates, velocities and
box size from the inpcrd file - Coordinates and velocities will be read
nstlim=250000000, ! Number of MD-steps to be performed
dt=0.002, ! The time step (psec). For temperatures above 300K, the
step size should be reduced.
ntc=2, ! Flag for SHAKE to perform bond length constraints -
bonds involving hydrogen are constrained
ntf=2, ! Force evaluation - bond interactions involving H-atoms
omitted (use with ntc=2)
cut=10, ! The nonbonded cutoff, in Angstroms
ntb=1, ! Impose periodic boundaries at constant volume
ntp=0, ! Flag for constant pressure dynamics - No pressure
ntt=3, ! Langevin thermostat
gamma_ln=2.0, ! Collision frequency in ps−1 for Langevin thermostat
ig=-1, ! Random seed for Langevin thermostat will be based on
the current date and time
temp0=300.0,! Reference temperature at which the system is to be kept, if
ntt > 0.
iwrap=1, ! The coordinates written to the restart and trajectory
files will be wrapped into a primary box
ntxo=2, ! Format of the final coordinates, velocities, and box
size written to the restart file - NetCDF file
ioutfm=1, ! The format of coordinate and velocity trajectory files
- Binary NetCDF trajectory
ntpr=10000, ! Print the progress of the run to output file every ntpr
ntwr=10000, ! Every ntwr steps during dynamics, the “restrt” file
will be written
ntwx=10000, ! Every ntwx steps, the coordinates will be written to
the mdcrd file
iamd=3, ! Accelerated MD - boost the whole potential with an
extra boost to the torsions

Launching the aMD with these parameters results in the following error:

cudaMemcpy GpuBuffer::Download failed an illegal memory access was

The error is being reproduced on different hardware (GTX 980 Ti and Tesla
K40s) and different units, while no error was previously observed on the
same units with other Amber simulations, including aMD simulations of other
proteins. Therefore I conclude that this is not a hardware issue.

HERE I FINALLY COME TO THE POINT. Thank you for reading this far.
When I "slightly" tune the ethreshp parameter and change it from
-343484.5895 to -413484.5895 the simulation starts smoothly and continues
to run.

1. Do you thinks the resulting simulation would still have a biological
significance after I`ve manually tuned this parameter?
2. Could this situation mean that my system is not properly relaxed
(although I`ve checked the density converges and RMSD plot goes flat)?

Thank you for your time.
AMBER mailing list
Received on Fri Aug 12 2016 - 14:00:02 PDT
Custom Search