GPUs on QHub#

Having access to GPUs is of prime importance for speeding up many computations by several orders of magnitude. QHub provides a way to achieve that, we will go through achieving that for each Cloud provider.

Clouds#

Google Cloud Platform#

By default the quota to spin up GPUs on GCP is 0. Make sure you have requested GCP Support to increase quota of allowed GPUs for your billing account to be the number of GPUs you need access to.

See GCP Pre-requisites here

Here are the changes needed in your qhub-config.yml file to get GPUs working with GCP:

1. Add GPU node group#

Add a node group for GPU instance in the node_groups section of google_cloud_platform section, and under the guest_accelerators section add the name of the GPU. A comprehensive list of GPU types can be found in at the Official GCP docs here: https://cloud.google.com/compute/docs/gpus

An example of getting GPUs on GCP:

google_cloud_platform:
  project: project-name
  region: us-central1
  zone: us-central1-c
  availability_zones:
  - us-central1-c
  kubernetes_version: 1.18.16-gke.502
  node_groups:
    # ....
    gpu-tesla-t4:
      instance: "n1-standard-8"     # 8 vCPUs, 30 GB: Skylake, Broadwell, Haswell, Sandy Bridge, and Ivy Bridge
      min_nodes: 0
      max_nodes: 5
      guest_accelerators:
        - name: nvidia-tesla-k80    # 12 GB GDDR5: Nividia Tesla K80
          count: 1

Notes:

One of the restrictions regarding GPUs on GCP is they can only be used with general-purpose N1 machine types, except A100 GPUs, which are only supported on a2 machine types
If you are not using the gcp provider in QHub but are using gcp (let’s say deploying on an existing gcp cluster). You will need to manually install NVIDIA drivers to the cluster - see documentation here.
See general limitations of GPUs on Google Kubernetes Engine.

2. Add GPU instance in the JupyterLab profiles#

profiles:
  jupyterlab:
# ....
  - display_name: Large GPU Instance
    description: Stable environment with 8 cpu / 32 GB ram and 1 Nvidia Tesla T4
    kubespawner_override:
      cpu_limit: 8
      cpu_guarantee: 7.25
      mem_limit: 32G
      mem_guarantee: 24G
      image: quansight/qhub-jupyterlab:v0.4.4
      extra_resource_limits:
        nvidia.com/gpu: 1
      node_selector:
          "cloud.google.com/gke-nodepool": "gpu-tesla-t4"

Amazon Web Services#

Here are the changes needed in your qhub-config.yml file to get GPUs working with AWS. Unlike GCP gpus are specified in the instance type this is due to how AWS exposes gpus to the user.

1. Add GPU node group#

amazon_web_services:
  region: us-west-2
#   ...
    gpu-g4:
      instance: g4dn.2xlarge     # NVIDIA Tesla T4
      min_nodes: 1
      max_nodes: 5
      gpu: true     # This marks the given instance type is gpu enabled.

2. Add GPU instance in the JupyterLab profiles#

profiles:
  jupyterlab:
# ....
    - display_name: GPU Instance
      description: Stable environment with 8 cpu / 32 GB ram and 1 Nvidia Tesla T4
      kubespawner_override:
        cpu_limit: 8
        cpu_guarantee: 7.25
        mem_limit: 32G
        mem_guarantee: 24G
        image: quansight/qhub-jupyterlab:v0.4.4
        extra_resource_limits:
          nvidia.com/gpu: 1
        node_selector:
          "eks.amazonaws.com/nodegroup": "gpu-g4"

Notes:

If you are not using the AWS provider in QHub but are using the AWS cloud (let’s say deploying on an existing AWS cluster), you will need to manually install NVIDIA drivers to the cluster. See documentation here.

DigitalOcean#

DigitalOcean does not support GPUs at the time of writing this.

Azure#

Azure does support GPUs in Kubernetes, but QHub doesn’t currently have official support for this.

Create conda environment to take advantage of GPUs#

First you need to consult the driver version of nvidia being used. This can easily be checked via the command nvidia-smi.

$ nvidia-smi
Thu May 20 18:05:14 2021
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.06    Driver Version: 450.51.06    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| 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  Tesla K80           Off  | 00000000:00:04.0 Off |                    0 |
| N/A   32C    P8    29W / 149W |      0MiB / 11441MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

The important section is CUDA Version. In general you should install a version of cudatoolkit that’s less than or equal to the cuda version (but not too old). If you install cudatoolkit-dev and cudatoolkit make sure that they are the same version exactly including minor version. Also in the near future cuda should have better ABI compatibility.

Below is an example gpu environment:

name: gpu-environment
channels:
 - conda-forge
dependencies:
 - python
 - cudatoolkit ==11.0.3
 - cudatoolkit-dev ==11.0.3
 - cupy
 - numba

We are working hard to make the GPU experience on Qhub as streamlined as possible. There are many small gotchas when working with GPUs and getting all the drivers installed properly.

Upgrade - Redeployment for Breaking Changes

Preemptible and Spot instances on QHub