Paddle Operator makes it easy to run paddle distributed training job on kubernetes by providing PaddleJob custom resource etc.
- Kubernetes >= 1.8
- kubectl
With kubernetes ready, you can install paddle operator with configuration in deploy folder (use deploy/v1 for kubernetes v1.16+ or deploy/v1beta1 for kubernetes 1.15-).
Create PaddleJob crd,
$ kubectl apply -f https://raw.githubusercontent.com/PaddleFlow/paddle-operator/main/deploy/v1/crd.yaml
A succeed creation leads to result as follows,
$ kubectl get crd
NAME CREATED AT
paddlejobs.batch.paddlepaddle.org 2021-02-08T07:43:24Z
Then deploy controller,
$ kubectl apply -f https://raw.githubusercontent.com/PaddleFlow/paddle-operator/main/deploy/v1/operator.yaml
the ready state of controller would be as follow,
$ kubectl -n paddle-system get pods
NAME READY STATUS RESTARTS AGE
paddle-controller-manager-698dd7b855-n65jr 1/1 Running 0 1m
By default, paddle controller runs in namespace paddle-system and only controll jobs in that namespace.
To run controller in a different namespace or controll jobs in other namespaces, you can edit charts/paddle-operator/values.yaml
and install the helm chart.
You can also edit kustomization files or edit deploy/v1/operator.yaml
directly for that purpose.
Deploy your first paddlejob demo with
$ kubectl -n paddle-system apply -f https://raw.githubusercontent.com/PaddleFlow/paddle-operator/main/deploy/examples/wide_and_deep.yaml
Check pods status
$ kubectl -n paddle-system get pods
Check paddle job status
$ kubectl -n paddle-system get pdj
Simply
$ kubectl delete -f https://raw.githubusercontent.com/PaddleFlow/paddle-operator/main/deploy/v1/crd.yaml -f https://raw.githubusercontent.com/PaddleFlow/paddle-operator/main/deploy/v1/operator.yaml
More configuration can be found in Makefile, clone this repo and enjoy it. If you have any questions or concerns about the usage, please do not hesitate to contact us.
Please refer to the 中文文档 for more information about paddle configuration.