The original script uses /lustre/atlas/proj-shared/csc249/sfw/NAMD_2.12_Linux-x86_64-multicore/namd2 without CUDA which can't seem to access at the minute. So I installed my own version of the binary from the NAMD website and got it working.

This led me to think it would be much better if we could use the GPU version of NAMD. So I downloaded that (NAMD_2.12_Linux-x86_64-multicore-CUDA) and tried using it but I get the error:

Pe 4 physical rank 4 binding to CUDA device 0 on nid11408: 'Tesla K20X'  Mem: 5759MB  Rev: 3.5
FATAL ERROR: CUDA error cudaMalloc(pp, sizeofT*len) in file src/CudaUtils.C, function allocate_device_T
 on Pe 4 (nid11408 device 0): all CUDA-capable devices are busy or unavailable

The command I was using in the namd.sh is:

$NAMD2 +ppn 7 +setcpuaffinity \
       +pemap 0,2,4,6,8,10,12 +commap 14 +idlepoll +devices 0 \
       $conf > $logfile 2>&1

I based this on a working test script for NAMD using aprun which uses:

aprun -n 1 -N 1 -d 8 $NAMD2 ++ppn 7 +setcpuaffinity +pemap 0,2,4,6,8,10,12 +commap 14 +idlepoll +devices 0 eq0.conf 2>&1

To run on one node.

Is there something obvious I am missing?