Submission for CUDA 5.0 to Tusker

README.md

In order to submit a previous version of CUDA to Tusker's GPU nodes, you need to bring along the CUDA library. Additionally, you need to append the LD_LIBRARY_PATH to include the current directory. In order to accomplish this, we used a wrapper script, run.sh, which sets the environment and runs the cuda code.

run.sh is set as the executable. The transfer input files includes the library (from the system path), and the main CUDA executable.

transfer_input_files = /util/opt/cuda/5.0/lib64/libcudart.so.5.0, main

The argument to run.sh is simply the executable to run, main.

NOTE: You will need to modify the transfer_input_files line to point to the location of the libcudart.so.5.0 on your system.

gpu.submit

universe = grid
 
grid_resource = condor tusker-gw1.unl.edu tusker-gw1.unl.edu:9619

executable = run.sh
arguments = ./main
transfer_input_files = /util/opt/cuda/5.0/lib64/libcudart.so.5.0, main

output = jm.out.condorc
error  = jm.err.condorc
log    = jm.log
WhenToTransferOutput = ON_EXIT
x509userproxy  = x509up_u1384
    
# Request the gpu
Request_GPUs = 1
queue

run.sh

#!/bin/sh

LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`pwd`
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`pwd`

exec "$@"