AresOperator - a multi-framework unified kubernetes operator

This repository implements a universal kubernetes operator for multiple AI training frameworks. It privides several machenisms such as state deducing and roles dependency handling, aiming at the general orchestration capability for AI workloads. See AresJob CRD definition.

The individually adapted frameworks are as follows.

• Tensorflow
• Pytorch
• MPI (for frameworks which support communiation via MPI, such as Horovod and Pytorch)
• Bagua
• Dask

We also provide the custom framework for not listed frameworks, and you can customize roles dependency and state deducing as required.

See samples for all frameworks.

Prerequisites

• Kubernetes
• Kubectl

Installation

Deploy the operator

Install Operator on an existing Kubernetes cluster.

git clone https://github.com/AresSys/AresOperator.git
cd AresOperator

# install crd
kubectl apply -f build/crd/.

# observer
kubectl apply -f build/observer.yaml

# operator
kubectl apply -f build/operator.yaml

Verify ares-operator is running

kubectl -n ares get pods

NAME                             READY   STATUS    RESTARTS   AGE
ares-observer-777b5c9cb7-59tv6   0/1     Running   0          17s
ares-observer-777b5c9cb7-qqfdn   0/1     Running   0          17s
ares-operator-55499bcb4c-sprq9   0/1     Running   0          10s

Examples

• MPI

# replace ${IMAGE} with a MPI image
kubectl apply -f config/samples/mpi-sample.yaml

Verify pods startup dependency

kubectl get pods

NAME                            READY   STATUS              RESTARTS   AGE
mpi-sample-launcher-0           0/1     Init:0/1            0          18s
mpi-sample-worker-0             0/1     ContainerCreating   0          18s

NOTE: See AresJobInitializer for the implementation of the initializer for AresOperator.