VolcanoJob
最近更新于 Jul 31, 2021
定义
Volcano Job,简称vcjob,是Volcano自定义的Job资源类型。区别于Kubernetes Job,vcjob提供了更多高级功能,如可指定调度器、支持最小运行pod数、 支持task、支持生命周期管理、支持指定队列、支持优先级调度等。Volcano Job更加适用于机器学习、大数据、科学计算等高性能计算场景。
样例
apiVersion: batch.volcano.sh/v1alpha1
kind: Job
metadata:
name: test-job
spec:
minAvailable: 3
schedulerName: volcano
priorityClassName: high-priority
policies:
- event: PodEvicted
action: RestartJob
plugins:
ssh: []
env: []
svc: []
maxRetry: 5
queue: default
volumes:
- mountPath: "/myinput"
- mountPath: "/myoutput"
volumeClaimName: "testvolumeclaimname"
volumeClaim:
accessModes: [ "ReadWriteOnce" ]
storageClassName: "my-storage-class"
resources:
requests:
storage: 1Gi
tasks:
- replicas: 6
name: "default-nginx"
template:
metadata:
name: web
spec:
containers:
- image: nginx
imagePullPolicy: IfNotPresent
name: nginx
resources:
requests:
cpu: "1"
restartPolicy: OnFailure
关键字段
- schedulerName
schedulerName表示该job的pod所使用的调度器,默认值为volcano,也可指定为default-scheduler。它也是tasks.template.spec.schedulerName的默认值。
- minAvailable
minAvailable表示运行该job所要运行的最少pod数量。只有当job中处于running状态的pod数量不小于minAvailable时,才认为该job运行正常。
- volumes
volumes表示该job的挂卷配置。volumes配置遵从kubernetes volumes配置要求。
- tasks.replicas
tasks.replicas表示某个task pod的副本数。
- tasks.template
tasks.template表示某个task pod的具体配置定义。
- tasks.policies
tasks.policies表示某个task的生命周期策略。
- policies
policies表示job中所有task的默认生命周期策略,在tasks.policies不配置时使用该策略。
- plugins
plugins表示该job在调度过程中使用的插件。
- queue
queue表示该job所属的队列。
- priorityClassName
priorityClassName表示该job优先级,在抢占调度和优先级排序中生效。
- maxRetry
maxRetry表示当该job可以进行的最大重启次数。
资源状态
- pending
pending表示job还在等待调度中,处于排队的状态。
- aborting
aborting表示job因为某种外界原因正处于中止状态,即将进入aborted状态。
- aborted
aborted表示job因为某种外界原因已处于中止状态。
- running
running表示job中至少有minAvailable个pod正在运行状态。
- restarting
restarting表示job正处于重启状态,正在中止当前的job实例并重新创建新的实例。
- completing
completing表示job中至少有minAvailable个数的task已经完成,该job正在进行最后的清理工作。
- completed
completing表示job中至少有minAvailable个数的task已经完成,该job已经完成了最后的清理工作。
- terminating
terminating表示job因为某种内部原因正处于终止状态,正在等到pod或task释放资源。
- terminated
terminated表示job因为某种内部原因已经处于终止状态,job没有达到预期就结束了。
- failed
failed表示job经过了maxRetry次重启,依然没有正常启动。
使用场景
- TensorFlow workload
以tensorflow为例,创建一个具有1个ps和2个worker的工作负载。
apiVersion: batch.volcano.sh/v1alpha1
kind: Job
metadata:
name: tensorflow-dist-mnist
spec:
minAvailable: 3 // 该job的3个pod必须都可用
schedulerName: volcano // 指定volcano为调度器
plugins:
env: []
svc: []
policies:
- event: PodEvicted // 当pod被驱逐时,重启该job
action: RestartJob
tasks:
- replicas: 1 // 指定1个ps pod
name: ps
template: // ps pod的具体定义
spec:
containers:
- command:
- sh
- -c
- |
PS_HOST=`cat /etc/volcano/ps.host | sed 's/$/&:2222/g' | sed 's/^/"/;s/$/"/' | tr "\n" ","`;
WORKER_HOST=`cat /etc/volcano/worker.host | sed 's/$/&:2222/g' | sed 's/^/"/;s/$/"/' | tr "\n" ","`;
export TF_CONFIG={\"cluster\":{\"ps\":[${PS_HOST}],\"worker\":[${WORKER_HOST}]},\"task\":{\"type\":\"ps\",\"index\":${VK_TASK_INDEX}},\"environment\":\"cloud\"};
python /var/tf_dist_mnist/dist_mnist.py
image: volcanosh/dist-mnist-tf-example:0.0.1
name: tensorflow
ports:
- containerPort: 2222
name: tfjob-port
resources: {}
restartPolicy: Never
- replicas: 2 // 指定2个worker pod
name: worker
policies:
- event: TaskCompleted // 2个worker完成任务时认为该job完成任务
action: CompleteJob
template: // worker pod的具体定义
spec:
containers:
- command:
- sh
- -c
- |
PS_HOST=`cat /etc/volcano/ps.host | sed 's/$/&:2222/g' | sed 's/^/"/;s/$/"/' | tr "\n" ","`;
WORKER_HOST=`cat /etc/volcano/worker.host | sed 's/$/&:2222/g' | sed 's/^/"/;s/$/"/' | tr "\n" ","`;
export TF_CONFIG={\"cluster\":{\"ps\":[${PS_HOST}],\"worker\":[${WORKER_HOST}]},\"task\":{\"type\":\"worker\",\"index\":${VK_TASK_INDEX}},\"environment\":\"cloud\"};
python /var/tf_dist_mnist/dist_mnist.py
image: volcanosh/dist-mnist-tf-example:0.0.1
name: tensorflow
ports:
- containerPort: 2222
name: tfjob-port
resources: {}
restartPolicy: Never
- argo workload
以argo为例,创建一个具有2个pod副本的工作负载,要求1个可用即可。
apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
generateName: volcano-step-job-
spec:
entrypoint: volcano-step-job
serviceAccountName: argo
templates:
- name: volcano-step-job
steps:
- - name: hello-1
template: hello-tmpl
arguments:
parameters: [{name: message, value: hello1}, {name: task, value: hello1}]
- - name: hello-2a
template: hello-tmpl
arguments:
parameters: [{name: message, value: hello2a}, {name: task, value: hello2a}]
- name: hello-2b
template: hello-tmpl
arguments:
parameters: [{name: message, value: hello2b}, {name: task, value: hello2b}]
- name: hello-tmpl
inputs:
parameters:
- name: message
- name: task
resource:
action: create
successCondition: status.state.phase = Completed
failureCondition: status.state.phase = Failed
manifest: | // Volcano Job的具体定义
apiVersion: batch.volcano.sh/v1alpha1
kind: Job
metadata:
generateName: step-job-{{inputs.parameters.task}}-
ownerReferences:
- apiVersion: argoproj.io/v1alpha1
blockOwnerDeletion: true
kind: Workflow
name: "{{workflow.name}}"
uid: "{{workflow.uid}}"
spec:
minAvailable: 1
schedulerName: volcano
policies:
- event: PodEvicted
action: RestartJob
plugins:
ssh: []
env: []
svc: []
maxRetry: 1
queue: default
tasks:
- replicas: 2
name: "default-hello"
template:
metadata:
name: helloworld
spec:
containers:
- image: docker/whalesay
imagePullPolicy: IfNotPresent
command: [cowsay]
args: ["{{inputs.parameters.message}}"]
name: hello
resources:
requests:
cpu: "100m"
restartPolicy: OnFailure
- MindSpore
以MindSpore为例,创建一个具有8个pod副本的工作负载,要求1个可用即可。
apiVersion: batch.volcano.sh/v1alpha1
kind: Job
metadata:
name: mindspore-cpu
spec:
minAvailable: 1
schedulerName: volcano
policies:
- event: PodEvicted
action: RestartJob
plugins:
ssh: []
env: []
svc: []
maxRetry: 5
queue: default
tasks:
- replicas: 8
name: "pod"
template:
spec:
containers:
- command: ["/bin/bash", "-c", "python /tmp/lenet.py"]
image: lyd911/mindspore-cpu-example:0.2.0
imagePullPolicy: IfNotPresent
name: mindspore-cpu-job
resources:
limits:
cpu: "1"
requests:
cpu: "1"
restartPolicy: OnFailure
说明事项
Volcano支持的计算框架
Volcano对当前主流的计算框架均能很好的支持,具体如下:
- TensorFlow
- Pytorch
- MindSpore
- PaddlePaddle
- Spark
- Flink
- OpenMPI
- Horovod
- MXNet
- Kubeflow
- Argo
- KubeGene
volcano和default-scheduler的选择
与default-scheduler相比,volcano在批处理方面进行了增强。它更适用于高性能计算场景,如机器学习、大数据应用和科学计算。