The shared library libcudnn.so.6 is needed in a version of tensorflow. Which can be added via

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/aaaaaa/cudnn/v6/cuda-8.0/lib64

Yet, locate command cannot find it after I have configured LD_LIBRARY_PATH. This is because the program uses a database of library files compiled by the system (and this only includes libraries from the global system library paths), so it doesn't actually look at LD_LIBRARY_PATH itself.

A solution (copied from email, courtesy of J.Field)

If you want the LD_LIBRARY_PATH setting to be automatically in place, you can either create a script that sets the variable before launching TensorFlow or you could add the setting to your .bashrc file in your

home directory. Either way, just add the

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/aaaaaa/cudnn/v6/cuda-8.0/lib64

line to the file (if you create a file to set the variable and call TensorFlow, remember to chmod +x to the file). Creating a file to launch TensorFlow is probably the better method as adding the LD_LIBRARY_PATH to .bashrc will make this setting global and every time you try to run a command, Linux will first look in that directory for any libraries and this will have a very slight performance penalty.

So I generated runBSS.sh, which contains