Enable access to GPUs#
GPUs are heavily used in machine learning workflows, and we support GPUs on all major cloud providers.
Setting up GPU nodes#
AWS#
Requesting Quota Increase#
On AWS, GPUs are provisioned by using P series nodes. Before they can be accessed, you need to ask AWS for increased quota of P series nodes.
Login to the AWS management console of the account the cluster is in.
Make sure you are in same region the cluster is in, by checking the region selector on the top right. This is very important, as getting a quota increase in the wrong region means we have to do this all over again.
Open the EC2 Service Quotas page
Select ‘Running On-Demand G and VT Instances’ quota - this provisions NVidia T4 GPUs (which are the
G4dn
instance type).Select ‘Request Quota Increase’.
Input the number of vCPUs needed. This translates to a total number of GPU nodes based on how many CPUs the nodes we want have. For example, if we are using G4 nodes with NVIDIA T4 GPUs, each
g4dn.xlarge
node gives us 1 GPU and 4 vCPUs, so a quota of 8 vCPUs will allow us to spawn 2 GPU nodes. We should fine tune this calculation for later, but for now, the recommendation is to give users a singleg4dn.xlarge
each, so the number of vCPUs requested should be4 * max number of GPU nodes
.Ask for the increase, and wait. This can take several working days, so do it as early as possible!
Setup GPU nodegroup on eksctl#
We use eksctl
with jsonnet
to provision our kubernetes clusters on
AWS, and we can configure a node group there to provide us GPUs.
In the
notebookNodes
definition in the appropriate.jsonnet
file, add a node definition for the appropriate GPU node type:{ instanceType: "g4dn.xlarge", tags+: { "k8s.io/cluster-autoscaler/node-template/resources/nvidia.com/gpu": "1" }, }
g4dn.xlarge
gives us 1 Nvidia T4 GPU and ~4 CPUs. Thetags
definition is necessary to let the autoscaler know that this nodegroup has 1 GPU per node. If you’re using a different machine type with more GPUs, adjust this definition accordingly.Render the
.jsonnet
file into a.yaml
file thateksctl
can useexport CLUSTER_NAME=<your_cluster>
jsonnet $CLUSTER_NAME.jsonnet > $CLUSTER_NAME.eksctl.yaml
Create the nodegroup
eksctl create nodegroup -f $CLUSTER_NAME.eksctl.yaml
This should create the nodegroup with 0 nodes in it, and the autoscaler should recognize this!
eksctl
will also setup the appropriate driver installer, so you won’t have to.
Setting up a GPU user profile#
Finally, we need to give users the option of using the GPU via a profile. This should be placed in the hub configuration:
jupyterhub:
singleuser:
extraEnv:
# Temporarily set for *all* pods, including pods without any GPUs,
# to work around https://github.com/2i2c-org/infrastructure/issues/1530
NVIDIA_DRIVER_CAPABILITIES: compute,utility
profileList:
- display_name: NVIDIA Tesla T4, ~16 GB, ~4 CPUs
description: "Start a container on a dedicated node with a GPU"
profile_options:
image:
display_name: Image
choices:
tensorflow:
display_name: Pangeo Tensorflow ML Notebook
slug: "tensorflow"
kubespawner_override:
image: "pangeo/ml-notebook:<tag>"
pytorch:
display_name: Pangeo PyTorch ML Notebook
default: true
slug: "pytorch"
kubespawner_override:
image: "pangeo/pytorch-notebook:<tag>"
kubespawner_override:
mem_limit: null
mem_guarantee: 14G
node_selector:
node.kubernetes.io/instance-type: g4dn.xlarge
extra_resource_limits:
nvidia.com/gpu: "1"
If using a
daskhub
, place this under thebasehub
key.The image used should have ML tools (pytorch, cuda, etc) installed. The recommendation is to provide Pangeo’s ml-notebook for tensorflow and pytorch-notebook for pytorch. We expose these as options so users can pick what they want to use.
Warning
Do not use the
latest
ormaster
tags - find a specific tag listed for the image you want, and use that.The
NVIDIA_DRIVER_CAPABILITIES
environment variable tells the GPU driver what kind of libraries and tools to inject into the container. Without setting this, GPUs can not be accessed.The
node_selector
makes sure that these user pods end up on the appropriate nodegroup we created earlier. Change the selector and themem_guarantee
if you are using a different kind of node
Do a deployment with this config, and then we can test to make sure this works!
Testing#
Login to the hub, and start a server with the GPU profile you just set up.
Open a terminal, and try running
nvidia-smi
. This should provide you output indicating that a GPU is present.Open a notebook, and run the following python code to see if tensorflow can access the GPUs:
import tensorflow as tf tf.config.list_physical_devices('GPU')
This should output something like:
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
If on an image with pytorch instead, try this:
import torch torch.cuda.is_available()
This should return
True
.Remember to explicitly shut down your server after testing, as GPU instances can get expensive!
If either of those tests fail, something is wrong and off you go debugging :)