Elastic Training Operator
Overview
Some distributed deep learning training framework like horovod support elastic training, which enables training job scale up and down the number of workers dynamically at runtime without interrupting the training process.
Et-operator provides a set of Kubernetes Custom Resource Definition that makes it easy to run horovod or AIACC elastic training in kubernetes. After submit a training job, you can scaleIn and scaleOut workers during training on demand, which can make your training job more elasticity and efficient.
Design
The et-operator, work with 3 new CRDs, TrainingJob, ScaleIn and ScaleOut.
TrainingJob
User submit a TrainingJob CR to specify a training job detail, like launcher's and worker's image, entrypoint command, replicas of workers.
The et-operator will receive the creation event, then create the sub resource (like pods, configmap, service, secret) of the TrainingJob, and
The TrainingJob will create workers pods and services, generate the Secret and ConfigMap for launcher pod,
when all workers ready, then operator will create the launcher pod and sync pods status.
After launcher pod exit, et-operator will uppdate TrainingJob phase to Success or Fail according to pod's exit code,
then do the cleanup.
ScaleIN
We can submit ScaleIn and ScaleOut resource to specify the scaleOut and scaleIn action of TrainingJob.
After the TrainingJob start running, et-operator will continuously check whether there are available ScaleIn and ScaleOut CR, and execute it.
In ScaleIn CR, we can specify the trainingJob's name and which workers that need to scaleIn (by count or detail worker's name).
When et-operator find an available ScaleIn CR, it will start to execute the scalein operation.
Firstly, it will update the host config of TrainingJob,
In horovod elastic mode, it needs a script that return the host's topology , the change of hosts will notify the launcher, then and it will shutdown the worker process not in hosts gracefully.
After the hostFile updated, et-operator start to detect whether the launch process exist,
when et-operator confirm that the scalein worker's launch process not exit, it will delete the worker's resource.
ScaleOut
In ScaleOut CR, we can specify the trainingJob's name and the count that we want to scaleout.
When et-operator start to execute the scalein operation,
different from scaleIn, it will firstly create the new worker's resources.
After worker's resources ready, then update the hostFile.
Setup
Installation
git clone http://github.com/aliyunContainerService/et-operator
cd et-operator
kubectl create -f config/deploy.yaml
Or you can customize some config, and run:
make deploy
You can check whether the Training Job custom resource is installed via:
kubectl get crd
NAME CREATED AT
scaleins.kai.alibabacloud.com 2020-11-11T11:16:13Z
scaleouts.kai.alibabacloud.com 2020-11-11T11:16:13Z
trainingjobs.kai.alibabacloud.com 2020-11-11T11:16:13Z
Check the operator status
kubectl -n kube-ai get pod
NAME READY STATUS RESTARTS AGE
et-operator-ddd56ff8c-tdr2n 1/1 Running 0 59s
User guide
Create a elastic training job
The training code need to be constructed in in elastic training mod, see detail. You can create an Training job by submit an TrainingJob YAML file. You can goto Horovod TrainingJob Example to see the example, and you can modify it in need.
kubectl apply -f examples/training_job.yaml
Check TrainingJob status
# kubectl get trainingjob
NAME PHASE AGE
elastic-training Running 77s
# kubectl get po
NAME READY STATUS RESTARTS AGE
elastic-training-launcher 1/1 Running 0 7s
elastic-training-worker-0 1/1 Running 0 10s
elastic-training-worker-1 1/1 Running 0 9s
ScaleIn training job
When you need to scaleIn the trainingJob workers, you can submit an ScaleIn CustomResource.
In Scalein Spec, you need to spec the name of TrainingJob, et-operator will find the match trainingJob and execute scaleIn to it. You can specify the workers to scaleIn ScaleIn by count or just specify the count ScaleIn by count .
kubectl create -f examples/scale_in_count.yaml
Check Scalein status
# kubectl get scalein
NAME PHASE AGE
scalein-sample-t8jxd ScaleSucceeded 11s
# kubectl get po
NAME READY STATUS RESTARTS AGE
elastic-training-launcher 1/1 Running 0 47s
elastic-training-worker-0 1/1 Running 0 50s
ScaleOut training job
When you need to scaleOut the trainingJob workers, you can submit an ScaleOut CustomResource, which just specify the count of workers you want to scaleOut.
kubectl create -f examples/scale_out.yaml
Check ScaleOut status
# kubectl get scaleout
NAME PHASE AGE
elastic-training-scaleout-9dtmw ScaleSucceeded 30s
# kubectl get po
NAME READY STATUS RESTARTS AGE
elastic-training-launcher 1/1 Running 0 2m5s
elastic-training-worker-0 1/1 Running 0 2m8s
elastic-training-worker-1 1/1 Running 0 40s
elastic-training-worker-2 1/1 Running 0 40s
Roadmap
- Use
kubectl execreplace ssh: the block major problem is thatkubectl execwill hang when target pod shutdown but what we want is to exit process. - Support spot instance in public cloud platform, before node released, we should trigger a scaleIn to the training worker who's workers on the spot nodes.
- Support fault tolerance
Developing
Prerequisites:
- Go >= 1.8
- kubebuilder >= 0.4.1
mkdir -p $(go env GOPATH)/src/github.com/aliyunContainerService
cd $(go env GOPATH)/src/github.com/aliyunContainerService
git clone https://github.com/aliyunContainerService/et-operator
cd et-operator
make
Build operator
export IMG=<image repo>
make docker-build
make docker-push
Running operator in local
make run-local