The following guide shows you how to spawn a GPU-instance in skyhigh. There are GPU's available within some of our openstack-clouds, and access are given by request (we do not have GPU's for all projects unfortunatley).
GPU Flavors
Currently we have the following GPU's available:
- At SkyHiGh (IIK's production instance):
- General purpose flavors, all IIK affiliates are eligible:
- dx5.8c90r.v100-8g: A flavor with 90GB RAM, 8 vCPU's and a 1/4 of a Tesla v100 (8GB GPU-RAM).
- de3.12c60r.a100-10g: A flavor with 60GB RAM, 12 vCPUs and 1/4 of a Tesla A100 (10GB GPU-RAM)
- Flavors only available for SFI NORCICS:
- de3.24c120r.a100-20g: A flavor with 120GB RAM, 24 vCPUs and 1/2 of a Tesla A100 (20GB GPU-RAM)
- de3.48c240r.a100-40g: A flavor with 240GB RAM, 48 vCPUS and 1/1 of a Tesla A100 (40GB GPU-RAM)
- Flavors only available for Norwegian Biomtetrics Lab:
- dx4.24c60r.p40-24g: A flavor with 60GB RAM, 24 vCPUs and 1/1 of a Tesla P40 (24GB GPU-RAM)
- de2.24c240r.a100-20g: A flavor with 240GB RAM, 24 vCPU's and 1/2 of a Tesla A100 (20GB GPU-RAM)
- de3.24c120r.a100d-20g: A flavor with 120GB RAM, 24 vCPU's and 1/4 of a Tesla A100 80GB (20GB GPU-RAM)
- dx4.24c60r.p40-24g: A flavor with 60GB RAM, 24 vCPUs and 1/1 of a Tesla P40 (24GB GPU-RAM)
- General purpose flavors, all IIK affiliates are eligible:
- At SkyLow (IIK's development instance):
- dx4.8c20r.m10-8G: A flavor with 20GB RAM, 8 vCPU's and one core of a Tesla M10 card (8GB GPU-RAM).
- dx4.24c110r.p100: A flavor with 110GB RAM, 24 vCPU's and a Tesla p100 card (16GB GPU-RAM)
- dx4.48c220r.2p100: A flavor with 220GB RAM, 48 vCPU's and two Tesla p100 cards (2*16GB GPU-RAM)
- At stackit (NTNU IT's production platform):
- dx4.28c120r.a100-20g: A flavor with 120GB RAM, 28vCPU's and 1/2 of a Tesla a100 (20GB GPU-RAM)
- dx5s.96c470r.a100d-80g.e3400g: A flavor with 470GB RAM, 96 vCPUs and a Tesla a100d (80 GB GPU-RAM) and 3.4TiB with compute-local flash storage.
- Only available for an IV-EPT project.
GPU-enabled images
We provide an image with pre-installed Nvidia driver and CUDA package. This image contains the word "GRID" in its name and are a regular Ubuntu Server LTS image with the following additions:
- There is a script which installs the correct version of the Nvidia GRID driver at boot. After a driver-update on the hypervisor this script will also update the driver in the VM.
- The CUDA-tools are pre-installed.
The installation of the drivers requires a restart; so the newly created instance will reboot shortly after creation.
Starting a GPU-instance
To start a GPU-enabled instance you simply create a VM as you usually would, but make sure to select:
- a GPU-enabled flavor
- a GRID-enabled image
After creating the VM it will boot, install the GRID-driver, and then reboot. So, do not be supprised if you suddenly loose access to your freshly booted machine. It will come back! 
It is vital that the VM uses the same driver as the hypervisor. If you update this driver your vGPU will cease to function. When the driver is updated on the hypervisor a new driver will also be installed in your VM.
Do not install nvtop. Its dependencies will break the driver.
CUDA-versions are tightly coupled with the driver version. The CUDA-version bundled with the image is proved working. You can therfore not expect that a newer CUDA-version would work.
cuDNN
Many of our GPU users will probably need Nvidia's cuDNN library. This is not pre-installed in our image, because Nvidia requires all users to register for the Nvidia Developer Program before dowloading. So, please follow the instructions here, to install it on your VM; and use tar file options. DO NOT USE THE DEB OR RPM ALTERNATIVE. Be sure to download the cuDNN version that corresponds to our current CUDA version.
Verify the GPU instance
After installation the presence of a GPU can be verified with lspci:
ubuntu@gputest:~$ lspci | grep NVIDIA 00:05.0 3D controller: NVIDIA Corporation Device 20f1 (rev a1)
You can also verify that a license for the GPU is acquired successfully (yes, we need licences to use our GPUs...):
ubuntu@gputest:~$ journalctl -u nvidia-gridd | tail Sep 04 08:54:23 jammy-gpu systemd[1]: Starting NVIDIA Grid Daemon... Sep 04 08:54:24 jammy-gpu nvidia-gridd[724]: Started (724) Sep 04 08:54:24 jammy-gpu systemd[1]: Started NVIDIA Grid Daemon. Sep 04 08:54:24 jammy-gpu nvidia-gridd[724]: Configuration parameter ( ServerAddress ) not set Sep 04 08:54:24 jammy-gpu nvidia-gridd[724]: vGPU Software package (0) Sep 04 08:54:24 jammy-gpu nvidia-gridd[724]: Ignore service provider and node-locked licensing Sep 04 08:54:24 jammy-gpu nvidia-gridd[724]: NLS initialized Sep 04 08:54:24 jammy-gpu nvidia-gridd[724]: Acquiring license. (Info: nvidiadls02.it.ntnu.no; NVIDIA Virtual Compute Server) Sep 04 08:54:27 jammy-gpu nvidia-gridd[724]: License acquired successfully. (Info: nvidiadls02.it.ntnu.no, NVIDIA Virtual Compute Server; Expiry: 2023-9-5 8:54:16 GMT)
The "nvidia-smi" tool will show you the GPU status
ubuntu@gputest:~$ nvidia-smi
Mon Sep 4 08:58:11 2023
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.125.06 Driver Version: 525.125.06 CUDA Version: 12.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 GRID V100D-8C On | 00000000:00:05.0 Off | N/A |
| N/A N/A P0 N/A / N/A | 0MiB / 8192MiB | 0% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
Starting from CUDA 12.x, the samples are no longer included in the install packages. If you want to verify that the GPU/license works, you have to download them from github:
$ git clone https://github.com/NVIDIA/cuda-samples.git
And the you can compile and run a sample. For instance you could do like so:
ubuntu@jammy-gpu:~$ cd cuda-samples/Samples/1_Utilities/deviceQuery
ubuntu@jammy-gpu:~/cuda-samples/Samples/1_Utilities/deviceQuery$ make
... lots-of-text-from-make ...
ubuntu@jammy-gpu:~/cuda-samples/Samples/1_Utilities/deviceQuery$ ./deviceQuery
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 1 CUDA Capable device(s)
Device 0: "GRID V100D-8C"
CUDA Driver Version / Runtime Version 12.0 / 12.0
CUDA Capability Major/Minor version number: 7.0
Total amount of global memory: 8192 MBytes (8589934592 bytes)
(080) Multiprocessors, (064) CUDA Cores/MP: 5120 CUDA Cores
GPU Max Clock rate: 1380 MHz (1.38 GHz)
Memory Clock rate: 877 Mhz
Memory Bus Width: 4096-bit
L2 Cache Size: 6291456 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total shared memory per multiprocessor: 98304 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 7 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device supports Managed Memory: No
Device supports Compute Preemption: Yes
Supports Cooperative Kernel Launch: Yes
Supports MultiDevice Co-op Kernel Launch: Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 0 / 5
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >
deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 12.0, CUDA Runtime Version = 12.0, NumDevs = 1
Result = PASS
If the output results in a "PASS" you should be fine. The GPU is then ready to use.
Using Docker
Our GPU-enabled VMs altso supports NVIDIA Container Toolkit in order to use docker with GPUs. Rule of thumb is to follow the current installation guide from NVIDIA. Don't bother with the driver pre-requsities. We have already sorted those out for you in our image. You will of course need to install docker before you begin. That part is also described in NVIDIA's documentation. We've summed up the commands below. The examples is for Ubuntu only.
Install docker
curl https://get.docker.com | sh && sudo systemctl --now enable docker
Install Nvidia Container Toolkit
# Enable the repositories distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \ && curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add - \ && curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list # Install the package sudo apt update && sudo apt -y install nvidia-docker2 # Restart the docker daemon sudo systemctl restart docker # Run a test to verifiy that it works sudo docker run --rm --gpus all nvidia/cuda:12.0.1-base-ubuntu22.04 nvidia-smi # Optionally run a test with Tensorflow that actually runs a bit of code on the GPU via docker sudo docker run --gpus all -it --rm tensorflow/tensorflow:2.14.0-gpu \ python -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"