doku:vsc3_gpu [VSC Wiki]

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
====== GPU computing and visualization =======

The following GPU devices are available:


^  Enterprise-grade Tesla k20m (kepler)  ^^
|  Total amount of global memory|4742 MBytes |
|  (13) Multiprocessors, (192) CUDA Cores/MP|2496 CUDA Cores |
|  GPU Clock rate|706 MHz |
|  Maximum number of threads per block|1024 |
|  Device has ECC support|Enabled |


^  Enterprise-grade Tesla m60 (maxwell)  ^^
|  Total amount of global memory|8114 MBytes |
|  (16) Multiprocessors, (128) CUDA Cores/MP|2048 CUDA Cores |
|  GPU Clock rate|1.18 GHz |
|  Maximum number of threads per block|1024 |
|  Device has ECC support|Disabled |


^  Consumer-grade GeForce GTX 1080 (pascal)  ^^
|  Total amount of global memory|8113 MBytes |
|  (20) Multiprocessors, (128) CUDA Cores/MP|2560 CUDA Cores |
|  GPU Clock rate|1.73 GHz |
|  Maximum number of threads per block|1024 |
|  Device has ECC support|Disabled |


  * Two nodes, n25-[005,006], with two Tesla k20m (kepler) GPUs each. Host systems are equipped with two Intel Xeon E5-2680 0 @ 2.70GHz each with 8 cores and 256GB of RAM.
  * One node, n25-007, with two Tesla m60 (maxwell) GPUs. n25-007 is equipped with 2 Intel Xeon E5-2650 v3 @ 2.30GHz, each with 10 cores and a host memory of 256GB RAM.
  * Ten nodes, n25-[011-020], with single GPUs of type GTX-1080 (pascal), where host systems are single socket 4-core Intel Xeon E5-1620 @ 3.5 GHz with 64GB RAM. 
  * Two shared-private nodes, n25-[021-022], each equipped with 8 GTX-1080 (pascal) devices hosted on dual socket 4-core Intel Xeon E5-2623 systems @ 2.6 GHz with 128GB RAM.
[[https://github.com/NVIDIA/gdrcopy|gdrdrv]] is loaded by default (also see [[https://devtalk.nvidia.com/default/topic/919381/gdrcpy-problem/|notes]] regarding gtx1080 cards).

-------
==== Slurm integration ====

There are several partions which include all identical compute nodes with the same amount of GPUs on each node:

<code>
gpu_gtx1080single   #  4 cpu cores, one gpu per node
gpu_gtx1080multi    # 16 cpu cores, eight gpus per node
gpu_k20m            # 16 cpu cores, two gpus per node
gpu_m60             # 16 cpu cores, one gpu per node
</code>

For each partition a identically named QOS is defined. Slurm usage is eg.:

<code>
#SBATCH -p gpu_gtx1080single
#SBATCH --qos gpu_gtx1080single
</code>


--------------------------------------------------

===== Visualization (!!currently not supported!!) ======

To make use of a gpu node for visualization you need to perform the following steps.

  - set a vnc password, this is needed when connecting to the vnc server, **this has to be done only once**: <code>module load TurboVNC/2.0.1
mkdir ${HOME}/.vnc
vncpasswd
Password: ******
Warning: password truncated to the length of 8.
Verify:   ******
Would you like to enter a view-only password (y/n)? n
</code>
  - allocate gpu nodes with this script: <code>sviz -a</code>
  - start vnc server: <code>sviz -r</code>
  - follow the instructions on the screen and connect from your local machine with a vncvier: <code>vncviewer -via <user>@vsc3.vsc.ac.at <node>::<port></code>

All options for sviz:
<code>
sviz -h
usage: /opt/sw/x86_64/generic/bin/sviz
Parameters:
-h      print this help
-a      allocate gpu nodes
-r      start vnc server on allocated nodes

options for allocating:
-t      set gpu type; default=gtx1080
-n      set gpu count; default=1

options for vnc server:
-g      set geometry; default=1920x1080


</code>





==== Linux/Windows - Tightvnc ====

On your local (Linux) workstation you can use any vnc client which supports a gateway parameter (usually, there is a ''-via'' option), e.g. TightVNC. You will be first asked for your VSC cluster password (and possibly your OTP if it has not been entered within the last 12 hours), and then for your VNC password which you entered in the previous step:
<code>user@localhost:~$ vncviewer -via user@vsc3.vsc.ac.at n25-001::5901
Password: 
Connected to RFB server, using protocol version 3.8
Enabling TightVNC protocol extensions
Performing standard VNC authentication
Password: 
Authentication successful
Desktop name "TurboVNC: n25-001:1 (user)"
VNC server default format:
  32 bits per pixel.
  Least significant byte first in each pixel.
  True colour: max red 255 green 255 blue 255, shift red 16 green 8 blue 0
Warning: Cannot convert string "-*-helvetica-bold-r-*-*-16-*-*-*-*-*-*-*" to type FontStruct
Using default colormap which is TrueColor.  Pixel format:
  32 bits per pixel.
  Least significant byte first in each pixel.
  True colour: max red 255 green 255 blue 255, shift red 16 green 8 blue 0
Tunneling active: preferring tight encoding</code>

You should now see a desktop like this:
{{:doku:tightvnc_screenshot.png?700|}}

Windows versions of TightVNC are also available.


==== OS X/Linux/Windows - TuboVNC ====

Under OS X it is suggested to use the [[http://sourceforge.net/projects/turbovnc/files/|TurboVNC]] client; but it may be used under Linux or Windows as well.
This is how you can setup the client connection to the VNC server:
  - Setup the connection: {{:doku:turbovnc:turbovnc_setup.png?600|}}
  - Enter your cluster password: {{:doku:turbovnc:turbovnc_clusterpw.png?300|}}
  - Enter your OTP: {{:doku:turbovnc:turbovnc_otp.png?300|}}
  - Enter your VNC password: {{:doku:turbovnc:turbovnc_vncpasswd.png?300|}}

A desktop will be displayed on your screen:
{{:doku:turbovnc:turbovnc_desktop.png|?650}}


==== VirtualGL ====

Load the module

<code>
module load VirtualGL/2.5.2
</code>


The following variables need to be set:

<code>
export VGL_DISPLAY=:0
export DISPLAY=:1
</code>

To make use of VirtualGL your application needs to be started with ''vglrun'':
<code>[user@n25-001 ~]$ vglrun <path_to_your_X_application></code>


---------------------
===== GPU computing =====

==== CUDA ====
Cuda toolkits are available in version 5.5, 7.5, 8.0.27 and 8.0.61 (which provide e.g. the nvcc compiler) and are accessible by loading the corresponding cuda module:
<code>module load cuda/5.5</code>
or
<code>module load cuda/7.5</code>
or
<code>module load cuda/8.0.27</code>
or
<code>module load cuda/8.0.61</code>

==== Batch jobs ====

To submit batch jobs, a sample job script ''my_gpu_job_script'' is:
<code>#!/bin/sh
#SBATCH -J gpucmp
#SBATCH -N 1
#SBATCH --partition=gpu
#SBATCH --qos=gpu_compute
#SBATCH --time=00:10:00
#SBATCH --gres=gpu:2
#SBATCH -C k20m

./my_code_which_runs_on_a_gpu
</code>
Submit the job:
<code>[user@l31 ~]$ sbatch my_gpu_job_script</code>


==== Controlling GPU utilization with nvidia-smi ====

The standard way to check whether an application actually makes use of GPUs and to what extent is by calling 
<code>
nvidia-smi
</code>
or better within a separate terminal
<code>
watch nvidia-smi
</code>
For further details please see the man page of ''nvidia-smi''

==== CUDA C References ====
{{:doku:cuda-docu:cuda_c_best_practices_guide.pdf|CUDA C best practices}}\\
{{:doku:cuda-docu:cuda_c_programming_guide.pdf|CUDA C programming guide}}\\
{{:doku:cuda-docu:cuda-gdb.pdf|CUDA GDB user manual}}\\

==== CUDA Libraries References ====
{{:doku:cuda-docu:cublas_library.pdf|CUBLAS library}}\\
{{:doku:cuda-docu:cufft_library.pdf|CUFFT library}}\\
{{:doku:cuda-docu:curand_library.pdf|CURAND library}}\\
{{:doku:cuda-docu:cusparse_library.pdf|CUSPARSE library}}\\

==== Additional Docu ====
More details found in dir ''...cuda/8.0.61/doc/pdf'' upon loading the corresponding cuda module.