Volcano is a batch scheduling system built on Kubernetes. It provides a suite of mechanisms (gang scheduling, job queues, fair scheduling policies) currently missing from Kubernetes that are commonly required by many classes of batch and elastic workloads. KubeRay's Volcano integration enables more efficient scheduling of Ray pods in multi-tenant Kubernetes environments.
Note that this is a new feature. Feedback and contributions welcome.
Volcano needs to be successfully installed in your Kubernetes cluster before enabling Volcano integration with KubeRay. Refer to the Quick Start Guide for Volcano installation instructions.
Deploy the KubeRay Operator with the --enable-batch-scheduler flag to enable Volcano batch scheduling support.
When installing via Helm, you can set the following in your values.yaml file:
batchScheduler:
enabled: true
Or on the command line:
# helm install kuberay-operator --set batchScheduler.enabled=true
Add the ray.io/scheduler-name: volcano label to your RayCluster CR to submit the cluster pods to Volcano for scheduling.
Example:
apiVersion: ray.io/v1alpha1
kind: RayCluster
metadata:
name: test-cluster
labels:
ray.io/scheduler-name: volcano
volcano.sh/queue-name: kuberay-test-queue
spec:
rayVersion: '2.4.0'
headGroupSpec:
rayStartParams: {}
replicas: 1
template:
spec:
containers:
- name: ray-head
image: rayproject/ray:2.4.0
resources:
limits:
cpu: "1"
memory: "2Gi"
requests:
cpu: "1"
memory: "2Gi"
workerGroupSpecs: []
The following labels can also be provided in the RayCluster metadata:
ray.io/priority-class-name: the cluster priority class as defined by Kubernetes here.volcano.sh/queue-name: the Volcano queue name the cluster will be submitted to.
If autoscaling is enabled, minReplicas will be used for gang scheduling, otherwise the desired replicas will be used.
In this example, we'll walk through how gang scheduling works with Volcano and KubeRay.
First, let's create a queue with a capacity of 4 CPUs and 6Gi of RAM:
$ kubectl create -f - <<EOF
apiVersion: scheduling.volcano.sh/v1beta1
kind: Queue
metadata:
name: kuberay-test-queue
spec:
weight: 1
capability:
cpu: 4
memory: 6Gi
EOF
The weight in the definition above indicates the relative weight of a queue in cluster resource division. This is useful in cases where the total capability of all the queues in your cluster exceeds the total available resources, forcing the queues to share among themselves. Queues with higher weight will be allocated a proportionally larger share of the total resources.
The capability is a hard constraint on the maximum resources the queue will support at any given time. It can be updated as needed to allow more or fewer workloads to run at a time.
Next we'll create a RayCluster with a head node (1 CPU + 2Gi of RAM) and two workers (1 CPU + 1Gi of RAM each), for a total of 3 CPU and 4Gi of RAM:
$ kubectl create -f - <<EOF
apiVersion: ray.io/v1alpha1
kind: RayCluster
metadata:
name: test-cluster-0
labels:
ray.io/scheduler-name: volcano
volcano.sh/queue-name: kuberay-test-queue
spec:
rayVersion: '2.4.0'
headGroupSpec:
rayStartParams: {}
replicas: 1
template:
spec:
containers:
- name: ray-head
image: rayproject/ray:2.4.0
resources:
limits:
cpu: "1"
memory: "2Gi"
requests:
cpu: "1"
memory: "2Gi"
workerGroupSpecs:
- groupName: worker
rayStartParams: {}
replicas: 2
minReplicas: 2
maxReplicas: 2
template:
spec:
containers:
- name: ray-head
image: rayproject/ray:2.4.0
resources:
limits:
cpu: "1"
memory: "1Gi"
requests:
cpu: "1"
memory: "1Gi"
EOF
Because our queue has a capacity of 4 CPU and 6Gi of RAM, this resource should schedule successfully without any issues. We can verify this by checking the status of our cluster's Volcano PodGroup to see that the phase is Running and the last status is Scheduled:
$ kubectl get podgroup ray-test-cluster-0-pg -o yaml
apiVersion: scheduling.volcano.sh/v1beta1
kind: PodGroup
metadata:
creationTimestamp: "2022-12-01T04:43:30Z"
generation: 2
name: ray-test-cluster-0-pg
namespace: test
ownerReferences:
- apiVersion: ray.io/v1alpha1
blockOwnerDeletion: true
controller: true
kind: RayCluster
name: test-cluster-0
uid: 7979b169-f0b0-42b7-8031-daef522d25cf
resourceVersion: "4427347"
uid: 78902d3d-b490-47eb-ba12-d6f8b721a579
spec:
minMember: 3
minResources:
cpu: "3"
memory: 4Gi
queue: kuberay-test-queue
status:
conditions:
- lastTransitionTime: "2022-12-01T04:43:31Z"
reason: tasks in gang are ready to be scheduled
status: "True"
transitionID: f89f3062-ebd7-486b-8763-18ccdba1d585
type: Scheduled
phase: Running
And checking the status of our queue to see that we have 1 running job:
$ kubectl get queue kuberay-test-queue -o yaml
apiVersion: scheduling.volcano.sh/v1beta1
kind: Queue
metadata:
creationTimestamp: "2022-12-01T04:43:21Z"
generation: 1
name: kuberay-test-queue
resourceVersion: "4427348"
uid: a6c4f9df-d58c-4da8-8a58-e01c93eca45a
spec:
capability:
cpu: 4
memory: 6Gi
reclaimable: true
weight: 1
status:
reservation: {}
running: 1
state: Open
Next we'll add an additional RayCluster with the same configuration of head / worker nodes, but a different name:
$ kubectl create -f - <<EOF
apiVersion: ray.io/v1alpha1
kind: RayCluster
metadata:
name: test-cluster-1
labels:
ray.io/scheduler-name: volcano
volcano.sh/queue-name: kuberay-test-queue
spec:
rayVersion: '2.4.0'
headGroupSpec:
rayStartParams: {}
replicas: 1
template:
spec:
containers:
- name: ray-head
image: rayproject/ray:2.4.0
resources:
limits:
cpu: "1"
memory: "2Gi"
requests:
cpu: "1"
memory: "2Gi"
workerGroupSpecs:
- groupName: worker
rayStartParams: {}
replicas: 2
minReplicas: 2
maxReplicas: 2
template:
spec:
containers:
- name: ray-head
image: rayproject/ray:2.4.0
resources:
limits:
cpu: "1"
memory: "1Gi"
requests:
cpu: "1"
memory: "1Gi"
EOF
Now check the status of its PodGroup to see that its phase is Pending and the last status is Unschedulable:
$ kubectl get podgroup ray-test-cluster-1-pg -o yaml
apiVersion: scheduling.volcano.sh/v1beta1
kind: PodGroup
metadata:
creationTimestamp: "2022-12-01T04:48:18Z"
generation: 2
name: ray-test-cluster-1-pg
namespace: test
ownerReferences:
- apiVersion: ray.io/v1alpha1
blockOwnerDeletion: true
controller: true
kind: RayCluster
name: test-cluster-1
uid: b3cf83dc-ef3a-4bb1-9c42-7d2a39c53358
resourceVersion: "4427976"
uid: 9087dd08-8f48-4592-a62e-21e9345b0872
spec:
minMember: 3
minResources:
cpu: "3"
memory: 4Gi
queue: kuberay-test-queue
status:
conditions:
- lastTransitionTime: "2022-12-01T04:48:19Z"
message: '3/3 tasks in gang unschedulable: pod group is not ready, 3 Pending,
3 minAvailable; Pending: 3 Undetermined'
reason: NotEnoughResources
status: "True"
transitionID: 3956b64f-fc52-4779-831e-d379648eecfc
type: Unschedulable
phase: Pending
Because our new cluster requires more CPU and RAM than our queue will allow, even though we could fit one of the pods with the remaining 1 CPU and 2Gi of RAM, none of the cluster's pods will be placed until there is enough room for all the pods. Without using Volcano for gang scheduling in this way, one of the pods would ordinarily be placed, leading to the cluster being partially allocated, and some jobs (like Horovod training) getting stuck waiting for resources to become available.
We can see the effect this has on scheduling the pods for our new RayCluster, which are listed as Pending:
$ kubectl get pods
NAME READY STATUS RESTARTS AGE
test-cluster-0-worker-worker-ddfbz 1/1 Running 0 7m
test-cluster-0-head-vst5j 1/1 Running 0 7m
test-cluster-0-worker-worker-57pc7 1/1 Running 0 6m59s
test-cluster-1-worker-worker-6tzf7 0/1 Pending 0 2m12s
test-cluster-1-head-6668q 0/1 Pending 0 2m12s
test-cluster-1-worker-worker-n5g8k 0/1 Pending 0 2m12s
If we dig into the pod details, we'll see that this is indeed because Volcano cannot schedule the gang:
$ kubectl describe pod test-cluster-1-head-6668q | tail -n 3
Type Reason Age From Message
---- ------ ---- ---- -------
Warning FailedScheduling 4m5s volcano 3/3 tasks in gang unschedulable: pod group is not ready, 3 Pending, 3 minAvailable; Pending: 3 Undetermined
Let's go ahead and delete the first RayCluster to clear up space in the queue:
$ kubectl delete raycluster test-cluster-0
The PodGroup for the second cluster has moved to the Running state, as there are now enough resources available to schedule the entire set of pods:
$ kubectl get podgroup ray-test-cluster-1-pg -o yaml
apiVersion: scheduling.volcano.sh/v1beta1
kind: PodGroup
metadata:
creationTimestamp: "2022-12-01T04:48:18Z"
generation: 9
name: ray-test-cluster-1-pg
namespace: test
ownerReferences:
- apiVersion: ray.io/v1alpha1
blockOwnerDeletion: true
controller: true
kind: RayCluster
name: test-cluster-1
uid: b3cf83dc-ef3a-4bb1-9c42-7d2a39c53358
resourceVersion: "4428864"
uid: 9087dd08-8f48-4592-a62e-21e9345b0872
spec:
minMember: 3
minResources:
cpu: "3"
memory: 4Gi
queue: kuberay-test-queue
status:
conditions:
- lastTransitionTime: "2022-12-01T04:54:04Z"
message: '3/3 tasks in gang unschedulable: pod group is not ready, 3 Pending,
3 minAvailable; Pending: 3 Undetermined'
reason: NotEnoughResources
status: "True"
transitionID: db90bbf0-6845-441b-8992-d0e85f78db77
type: Unschedulable
- lastTransitionTime: "2022-12-01T04:55:10Z"
reason: tasks in gang are ready to be scheduled
status: "True"
transitionID: 72bbf1b3-d501-4528-a59d-479504f3eaf5
type: Scheduled
phase: Running
running: 2
Checking the pods again, we see that the second cluster is now up and running:
$ kubectl get pods
NAME READY STATUS RESTARTS AGE
test-cluster-1-worker-worker-n5g8k 1/1 Running 0 9m4s
test-cluster-1-head-6668q 1/1 Running 0 9m4s
test-cluster-1-worker-worker-6tzf7 1/1 Running 0 9m4s
Finally, we'll cleanup the remaining cluster and queue:
$ kubectl delete raycluster test-cluster-1
$ kubectl delete queue kuberay-test-queue
Reach out to @tgaddair for questions regarding usage of this integration.