PGI GPU/CUDA Programming on the Princeton adroit cluster

o PGI CUDA Fortran Manual, examples, and overviews

  /opt/pgi/linux86-64/9.0/doc/pgicudaforug.pdf 
  http://www.pgroup.com/lit/samples/pgi_accelerator_examples.tar
  http://www.pgroup.com/lit/articles/insider/v1n1a1.htm  
  http://www.pgroup.com/lit/articles/insider/v1n2a1.htm

o The adroit cluster 

  Portal, adroit.princeton.edu, can be ssh-ed from PPPL portals.
  Apply for access through http://www.princeton.edu/~cses/beowulf
  
  Support is available from Computational Science and Engineering
  via e-mail address cses@princeton.edu.  

o Using the GPU software on adroit-001 and adroit-002

  On adroit-001 or adroit-002, you can run:
    $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib
    $ /usr/local/NVIDIA_CUDA_SDK/C/bin/linux/release/deviceQuery
  to get info on the Tesla itself.

  Through PBS, you need to give the "tesla" attribute:
    #PBS -l nodes=2:ppn=8:tesla,walltime=10:00:00

o Building a program on adroit-001 or adroit-002

  $ module load pgi/9.0/64 
  $ NVIDIA=/usr/local/cuda
  $ export PATH=$NVIDIA/bin:$PATH
  $ export LD_LIBRARY_PATH=$NVIDIA/lib64:$LD_LIBRARY_PATH

  $ pgfortran -o f1.exe f1.f90 -ta=nvidia -Minfo=accel -fast
  $ ./f1.exe

Updated: November 6, 2009