The Stadius sista-nc-3 server has a Tesla K40 GPU donated by NVIDIA that can be used by researchers in the Stadius research group. This Github page will explain how to use this GPU via libraries such as Theano, and will give you pointers where and how to learn more about GPU programming.
If you use this GPU for your research you should acknowledge the NVIDIA corporation by mentioning "The Tesla K40 used for this research was donated by the NVIDIA Corporation.".
GPU-accelerated computing is the use of a graphics processing unit (GPU) to accelerate certain computations. GPUs can gain speedups over a CPU because of their highly parallel architecture. The following picture illustrates this difference in parallelization. While a CPU only has a few computational cores, the CPU typically has a few hundred.
While a single GPU core is typically much more limited in what it can do and is much slower than a CPU core, a usage of many cores in parallel can still give certain speedups over sequential CPU programs. The speedup that can be gained is depended on the problem you are trying to solve, if your problem can be parallelized easily your speedups can be huge. If you problem is almost completely sequential, almost no speedups can be made, and your program might even run slower. Typical linear algebra equations are a good candidate to parallelize because of the parallel nature of vector additions and multiplications.
To set up this tutorial and copy the files you need to your home directory on the server you can follow the following steps.
Windows users can download Cygwin together with SSH to use the bash command line functionality on Windows (make sure to select the SSH package during installation).
Login to the sista-nc-3 server with your Stadius account by executing the following command in your shell (be sure that SSH is installed and configured, if you don't know how to do this, ask your nearest colleague):
ssh -A -t <user_name>@ssh.esat.kuleuven.be ssh -A -t sista-nc-3
With <user_name> your user name. This might ask you password corresponding to your user name.
Then download this Github repository via git by running the following command on the server side terminal (the one you just logged into):
git clone https://github.com/peterroelants/Stadius_GPU.git
The ssh -A -t <user_name>@ssh.esat.kuleuven.be ssh -A -t sista-nc-3 command will create a multi-hop SSH connection to sista-nc-3 by first going through the ssh.esat.kuleuven.be login server. The -A option forwards your user name to sista-nc-3 via ssh.esat.kuleuven.be without needing it again. The -t option forces to open a terminal on sista-nc-3. The result of this command will be the same as first running ssh <user_name>@ssh.esat.kuleuven.be and then ssh <user_name>@sista-nc-3.
git clone https://github.com/peterroelants/Stadius_GPU.git Uses Git's' clone command to download the necessary files in https://github.com/peterroelants/Stadius_GPU.
A guide to setup and use Python with the GPU is provided in this link.
This tutorial explains how to install Anaconda Python and Theano in your home directory and set up an IPython notebook so you can run your programs from the browser on your local machine.
In Stadius the Matlab version comes with the Parallel Computing Toolbox. This toolbox allows to use the GPU if there is one available, and since the sista-nc-3 server has one available we can make use of it.
A guide to use Matlab with the GPU is provided in this link.
To learn more about GPU programming check out the following 2 online courses: