cloudwise 0.1

Set up a Kubernetes cluster for distributed AI research

Cloudwise

Installation

Usage

FAQs

---

cloudwise is Surreal’s cloud infrastructure provisioner based on Terraform. Surreal’s website and github.

It prepares a kubernetes cluster using terraform. It generates .tf.json files that are also recognized by Symphony.

Installation

• Cloud wise runs in python 3

• Do git clone [e-mail]:SurrealAI/cloudwise.git && cd cloudwise

• Run pip install -e . in this directory.

• Install terraform following instructions here

• Install kubectl following instructions here

Usage

• (Optional, Recommended) Create and work in a clean directory as running terraform would generate relevant files.

> mkdir surreal
> cd surreal

Google Cloud

• You first need to setup credentials for terraform to access google cloud. See guide here. Choose one of the two methods:

  • Run the following command

  gcloud auth application-default login

  or

  • Go to the api key management page https://console.cloud.google.com/apis/credentials/serviceaccountkey and select Create new service account. You would need to give the service account sufficient permissions to do things properly. Project editor would suffice but is also more than enough. You can then generate and download the key, (json format is fine). Put the path to the .json file into the commandline argument when prompted.

• Follow the instructions in the commandline tool.

> cloudwise-gke

It will provide instructions and generate a <cluster_name>.tf.json file which terraform recognizes. If you have generated a .json credential file, you should provide it when prompted. * terraform init && terraform plan describes changes to be made. * terraform apply makes the changes to your cloud project. * After cluster creation, obtain credentials for kubectl.

> gcloud container clusters get-credentials <cluster_name>

• If you have GPUs in your cluster, create the daemon set to install drivers, see documentation.

> kubectl apply -f https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/stable/nvidia-driver-installer/cos/daemonset-preloaded.yaml

• The generated <cluster_name>.tf.json is also recognized by Symphony’s scheduling mechanism and Surreal. So you may want to link to it

• If you want to remove everything, run terraform destroy

AWS

Stay tuned

Azure

Stay tuned

FAQs:

• Terraform install fails.

  • If you are seeing error: ... API has not been used in project...: during terraform apply, go to the Kubernetes Engine tab and/or Compute Engine tab on your google cloud console to enable their APIs.

• GPU nodes are not scaling up.

  • Check if the driver installation daemon set is running (see documentation).