Computational Drug Discovery and Design

MDlast=$Njobs

MDcurr=$MDfirst

MDinp=0

while [ $MDcurr -le $MDlast ]

do

echo -n "Job $MDcurr started on "

date

let "MDinp = ${MDcurr} - 1"

$AMBERHOME/bin/pmemd.cuda -O -i mdin$MDcurr \

-o $BASE$MDcurr.out \

-p $BASE.prmtop \

-c $BASE$MDinp.rst \

-ref $BASE$MDinp.rst \

-r $BASE$MDcurr.rst \

-inf mdinfo$MDcurr \

-x $BASE$MDcurr.mdcrd

let "MDcurr = ${MDcurr} + 1"

done

echo -n "Jobs finished on "

date

The deviceQuery command comes with the CUDA Toolkit

(https://developer.nvidia.com/cuda-toolkit), which has to be

installed before using any GPU-related software. The CUDA ver-

sion should correspond to the one with which the software was

developed. If multiple GPU cards are to be used in parallel, the list

of selected devices (variable “GPU” in the script above) contains

comma-separated identifiers (e.g., “0,1”), and the first line of the

launch command should be:

mpirun -n NGPUS $AMBERHOME/bin/pmemd.cuda.MPI -O -i mdin

$MDcurr \

where NGPUS is the number of GPU cards supposed to work in

parallel on a given task. In order to adjust the input MD files to the

above script, rename the file sys.inpcrd (seeSubheading3.2,step 1)

to sys0.rst before the start of the simulations (seeNote 6).

2. Log files from each stage of the simulation should be inspected
   for possible problems. If the first MD stage fails, it may be due
   to the failure of the first two minimization steps to remove the
   hot spots in the initial structure. Consequently, the input
   structure has to be examined and corrected. Ideally, the simu-
   lation would terminate without problems. However, large

156 Gre ́gory Menchon et al.