Week 12 (April 2)

Tuesday

• Tips on homework; discussion of projects

Accessing GPUs for programming in CUDA.

We will be working on the Redhawk cluster to access GPUs.

Thursday

Logging into Redhawk

ssh <yourname>@redhawk.boisestate.edu

• Set up an ssh key to avoid being asked to enter your password each time.

• Create a .bashrc file.

Here is a minimal example that you should find useful.

# .bashrc file   
if [ -f /etc/bashrc ]; then
	. /etc/bashrc
fi

# User specific aliases and functions
module load gcc/4.8.1
module load slurm 
module load openmpi/gcc-4.8.1/cuda75/1.10.1
module load cuda75/toolkit/7.5
module load python/3.5.1
# source ~/.bash_alias
export PATH=.:${PATH}

In this file, there are commands to

* Load the CUDA toolkit which is needed to compile CUDA programs.  (Note: version 8.0 is not compatible with the installed driver;  so we only load version 7.5). 
* Load OpenMPI (an MPI implementation; alternative to MPICH)
* Load the version of gcc needed to compile CUDA ( version ~ 4.8).  

Here is a simple .bash_alias file:

# .bash_alias
alias ls="ls -F --color"
alias rm="rm -i"
alias cp="cp -i"

Introduction to programming in CUDA

A CUDA program looks very much like a C program, but with some strange "kernel" commands. Here is a very simple Hello, World program that calls the CUDA kernel

#include <stdio.h>

__global__ void kernel( void ) 
{
    /* Do something fun! */
}

int main(void) 
{
    kernel<<<1,1>>>();        /* Call a CUDA kernel */
    printf( "Hello, World!\n" );
}

We create this file and save it with an extension .cu to indicate that it should be compiled as a CUDA file. To compile this, we use the NVIDIA compiler nvcc:

$ nvcc -o simple_kernel simple_kernel.cu

There are many simple examples like this on the Nvidia Cuda by Example webpage. These examples are explained in more detail in the book Cuda By Example (Sanders and Kandrot).