Dynamic Resource Oversubscription

Features

After dynamic resource oversubscription and elastic scaling are enabled in a node pool, oversubscribed resources change rapidly because the resource usage of high-priority applications changes in real time. To prevent frequent node scale-ins and scale-outs, do not consider oversubscribed resources when evaluating node scale-ins.

Hybrid deployment is supported, and CPU and memory resources can be oversubscribed. The key features are as follows:

• Offline jobs preferentially run on oversubscribed nodes.

  If both oversubscribed and non-oversubscribed nodes exist, the former will score higher than the latter and offline jobs are preferentially scheduled to oversubscribed nodes.

• Online jobs can use only non-oversubscribed resources if scheduled to an oversubscribed node.

  Offline jobs can use both oversubscribed and non-oversubscribed resources of an oversubscribed node.

• In the same scheduling period, online jobs take precedence over offline jobs.

  If both online and offline jobs exist, online jobs are scheduled first. When the node resource usage exceeds the upper limit and the node requests exceed 100%, offline jobs will be evicted.

• CPU/Memory resources can be isolation by kernel.

  CPU isolation: Online jobs can quickly preempt CPU resources of offline jobs and suppress the CPU usage of the offline jobs.

  Memory isolation: When system memory resources are used up and OOM Kill is triggered, the kernel evicts offline jobs first.

• kubelet offline jobs obey the following admission rules:

  After the pod is scheduled to a node, kubelet starts the pod only when the node resources can meet the pod request (predicateAdmitHandler.Admit). kubelet starts the pod when both of the following conditions are met:

  • The total request of pods to be started and online running jobs < allocatable nodes
  • The total request of pods to be started and online/offline running job < allocatable nodes+oversubscribed nodes
• Resource oversubscription and hybrid deployment can be configured separately.

  Enabling hybrid deployment of a node pool also enables oversubscription by default. Nodes are then labeled with both volcano.sh/colocation="true" and volcano.sh/oversubscription="true". To use hybrid deployment for both online and offline jobs without oversubscription, simply disable oversubscription in hybrid deployment settings. This will remove the volcano.sh/oversubscription="true" label.

  The following table lists the features that can be used after hybrid deployment or oversubscription is enabled.

  Hybrid Deployment	Oversubscription	Oversubscription Resource	Scenario for Evicting Offline Pods
  No	No	No	No
  Yes	No	No	The actual resource usage of a node exceeds the upper limit.
  No	Yes	Yes	The actual resource usage of a node exceeds the upper limit and the pod requests on the node exceed 100%.
  Yes	Yes	Yes	The actual resource usage of a node exceeds the upper limit.

How to Use

Consider the cluster version when determining how to make use of resource oversubscription. For details, see Table 1.

When cloud native hybrid deployment is enabled, resource oversubscription is enabled by default. For details, see Resource Oversubscription in Cloud Native Hybrid Deployment. The cloud native hybrid deployment add-on, volcano-agent, reports oversubscribed resources and evicts service load from nodes. Its core features include CPU/memory suppression, dynamic resource oversubscription, CPU burst, and hierarchical QoS control of egress.

To maintain compatibility with earlier versions, the function of manually setting kubelet parameters to enable resource oversubscription is still available. For details, see Compatible kubelet Oversubscription (Not Recommended). However, this solution only provides basic functions and has not been updated. It does not support new features like CPU burst and hierarchical QoS control of egress.

Resource Oversubscription in Cloud Native Hybrid Deployment

Specifications

• Cluster version
  • v1.23: v1.23.9-r0 or later
  • v1.25: v1.25.4-r0 or later
• Cluster type: CCE Standard or CCE Turbo
• Node OS: Huawei Cloud EulerOS 2.0
• Node type: ECS on x86
• Volcano version: 1.10.0 or later

Constraints

• Before enabling oversubscription, ensure that the overcommit add-on is not enabled on Volcano.
• Running pods cannot be converted between online and offline services. To convert services, rebuild pods.
• If you have set cpu-manager-policy to statically bind CPU cores on a node, do not assign the QoS class of Guaranteed to offline pods. This is because offline pods may occupy the CPUs of online pods, leading to an online pod startup failure and offline pods failing to start even though they have been successfully scheduled. To prevent this, switch the pods to online pods if CPU core binding is required.
• If cpu-manager-policy is set to static CPU core binding on a node, do not bind CPU cores to all online pods. This is because doing so can cause online pods to occupy all available CPU or memory resources, leaving only a small number of oversubscribed resources.

1. Log in to the CCE console and click the cluster name to access the cluster console.
2. In the navigation pane, choose Nodes. On the Node Pools tab page, locate the target node pool and choose More > Mixed configuration.

   Ensure that node pool hybrid deployment and resource oversubscription are enabled. For details, see Procedure.

   

3. (Optional) Adjust resource oversubscription parameters.

   Table 2 Resource oversubscription parameters

   Parameter	Description
   High CPU Eviction Threshold (%)	When the CPU usage of a node exceeds the specified value, offline job eviction is triggered and the node becomes unschedulable. The default value is 80, indicating that offline job eviction is triggered when the CPU usage of a node exceeds 80%.
   Low CPU Eviction Threshold (%)	When the CPU usage of a node is higher than the upper limit, offline jobs will be evicted. The node accepts the offline jobs again only when the CPU usage of the node is lower than the lower limit. The default value is 30, indicating that offline jobs are accepted again when the CPU usage of a node is lower than 30%.
   High Memory Eviction Threshold (%)	When the memory usage of a node exceeds the specified value, offline job eviction is triggered and the node becomes unschedulable. The default value is 60, indicating that offline job eviction is triggered when the memory usage of a node exceeds 60%.
   Low Memory Eviction Threshold (%)	When the memory usage of a node is higher than the upper limit, offline jobs will be evicted. The node accepts the offline jobs again only when the memory usage of the node is lower than the lower limit. The default value is 30, indicating that offline jobs are accepted again when the memory usage of a node is less than 30%.

4. Check the Volcano configuration.

   kubectl edit cm volcano-scheduler-configmap -n kube-system

   Check the oversubscription configuration in volcano-scheduler-configmap. Ensure that the add-on configuration does not contain the overcommit add-on. If - name: overcommit exists, delete this configuration.

   ...
   data:
     volcano-scheduler.conf: |
       actions: "allocate, backfill, preempt"   # Configure a preemption action.
       tiers:
       - plugins:
         - name: gang
           enablePreemptable: false
           enableJobStarving: false
         - name: priority
         - name: conformance
         - name: oversubscription
       - plugins:
         - name: drf
         - name: predicates
         - name: nodeorder
         - name: binpack
       - plugins:
         - name: cce-gpu-topology-predicate
         - name: cce-gpu-topology-priority
         - name: cce-gpu
   ...

5. Create resources at a high- and low-priorityClass, respectively.

   cat <<EOF | kubectl apply -f -
   
   apiVersion: scheduling.k8s.io/v1
   description: Used for high priority pods
   kind: PriorityClass
   metadata:
     name: volcano-production
   preemptionPolicy: PreemptLowerPriority
   value: 999999
   ---
   apiVersion: scheduling.k8s.io/v1
   description: Used for low priority pods
   kind: PriorityClass
   metadata:
     name: volcano-free
   preemptionPolicy: PreemptLowerPriority
   value: -90000
   
   EOF

6. Deploy online and offline jobs.

   For both online and offline jobs, set schedulerName to volcano to enable Volcano.

   • If you enable low-priority services when creating an offline job workload, the annotation volcano.sh/qos-level: "-1" will be added to it.

     

     Set priorityClassName to volcano-free.

     kind: Deployment 
     apiVersion: apps/v1 
     spec: 
       replicas: 4 
       template: 
         metadata: 
           annotations: 
             metrics.alpha.kubernetes.io/custom-endpoints: '[{"api":"","path":"","port":"","names":""}]' 
             volcano.sh/qos-level: "-1" # Offline job annotation
         spec: 
           schedulerName: volcano             # Volcano is used.
           priorityClassName: volcano-free         # volcano-free priorityClass
           ...

   • For online jobs, set priorityClassName to volcano-production.

     kind: Deployment 
     apiVersion: apps/v1 
     spec: 
       replicas: 4 
       template: 
         metadata: 
           annotations: 
             metrics.alpha.kubernetes.io/custom-endpoints: '[{"api":"","path":"","port":"","names":""}]' 
         spec: 
           schedulerName: volcano          # Volcano is used.
           priorityClassName: volcano-production   # volcano-production priorityClass
           ...

7. Run the following command to check the number of oversubscribed resources and the resource usage:

   kubectl describe node <nodeIP>
   # kubectl describe node 192.168.0.0 
   Name:             192.168.0.0 
   Roles:              <none> 
   Labels:           ... 
                     volcano.sh/oversubscription=true 
   Annotations:      ... 
                     volcano.sh/oversubscription-cpu: 2335 
                     volcano.sh/oversubscription-memory: 341753856 
   Allocatable: 
     cpu:               3920m 
     memory:            6263988Ki 
   Allocated resources: 
     (Total limits may be over 100 percent, i.e., overcommitted.) 
     Resource           Requests      Limits 
     --------           --------      ------ 
     cpu                 4950m (126%)  4950m (126%) 
     memory             1712Mi (27%)  1712Mi (27%)

Compatible kubelet Oversubscription (Not Recommended)

Specifications

• Cluster version
  • v1.19: v1.19.16-r4 or later
  • v1.21: v1.21.7-r0 or later
  • v1.23: v1.23.5-r0 or later
  • v1.25 or later
• Cluster type: CCE Standard or CCE Turbo
• Node OS: EulerOS 2.9 (kernel-4.18.0-147.5.1.6.h729.6.eulerosv2r9.x86_64) or Huawei Cloud EulerOS 2.0
• Node type: ECS
• Volcano version: 1.7.0 or later

Constraints

• Before enabling oversubscription, ensure that the overcommit add-on is not enabled on Volcano.
• Modifying the label of an oversubscribed node does not affect the running pods.
• Running pods cannot be converted between online and offline services. To convert services, rebuild pods.
• If the label volcano.sh/oversubscription=true is configured for a node in the cluster, the oversubscription configuration must be added to the Volcano add-on. Otherwise, the scheduling of oversold nodes will be abnormal. Ensure that you have correctly configure labels because the scheduler does not check the add-on and node configurations. For details, see Table 3.
• To disable oversubscription, perform the following operations:
  • Remove the volcano.sh/oversubscription label from the oversubscribed node.
  • Set over-subscription-resource to false.
  • Modify the configmap of Volcano Scheduler named volcano-scheduler-configmap and remove the oversubscription add-on.
• If you have set cpu-manager-policy to statically bind CPU cores on a node, do not assign the QoS class of Guaranteed to offline pods. This is because offline pods may occupy the CPUs of online pods, leading to an online pod startup failure and offline pods failing to start even though they have been successfully scheduled. To prevent this, switch the pods to online pods if CPU core binding is required.
• If cpu-manager-policy is set to static CPU core binding on a node, do not bind CPU cores to all online pods. This is because doing so can cause online pods to occupy all available CPU or memory resources, leaving only a small number of oversubscribed resources.

Deployment Example

The following uses an example to describe how to deploy online and offline jobs in hybrid mode.

1. Configure a cluster with two nodes, one oversubscribed and the other non-oversubscribed.

   # kubectl get node
   NAME           STATUS   ROLES    AGE    VERSION
   192.168.0.173   Ready    <none>   4h58m   v1.19.16-r2-CCE22.5.1
   192.168.0.3     Ready    <none>   148m    v1.19.16-r2-CCE22.5.1

   • 192.168.0.173 is an oversubscribed node (with the volcano.sh/oversubscription=true label).
   • 192.168.0.3 is a non-oversubscribed node (without the volcano.sh/oversubscription=true label).

   # kubectl describe node 192.168.0.173
   Name:               192.168.0.173
   Roles:              <none>
   Labels:             beta.kubernetes.io/arch=amd64
                       ...
                       volcano.sh/oversubscription=true

2. Submit offline job creation requests. If resources are sufficient, all offline jobs will be scheduled to the oversubscribed node.
   The offline job template is as follows:

   apiVersion: apps/v1
   kind: Deployment
   metadata:
     name: offline
     namespace: default
   spec:
     replicas: 2
     selector:
       matchLabels:
         app: offline
     template:
       metadata:
         labels:
           app: offline
         annotations:
           volcano.sh/qos-level: "-1"       # Offline job label
       spec:
         schedulerName: volcano             # Volcano is used.
         priorityClassName: volcano-free         # volcano-free priorityClass
         containers:
           - name: container-1
             image: nginx:latest
             imagePullPolicy: IfNotPresent
             resources:
               requests:
                 cpu: 500m
                 memory: 512Mi
               limits:
                 cpu: "1"
                 memory: 512Mi
         imagePullSecrets:
           - name: default-secret

   Offline jobs are scheduled to the oversubscribed node.

   # kubectl get pod -o wide
   NAME                      READY   STATUS   RESTARTS  AGE     IP             NODE 
   offline-69cdd49bf4-pmjp8   1/1    Running   0         5s    192.168.10.178   192.168.0.173
   offline-69cdd49bf4-z8kxh   1/1    Running   0         5s    192.168.10.131   192.168.0.173

3. Submit online job creation requests. If resources are sufficient, the online jobs will be scheduled to the non-oversubscribed node.
   The online job template is as follows:

   apiVersion: apps/v1
   kind: Deployment
   metadata:
     name: online
     namespace: default
   spec:
     replicas: 2
     selector:
       matchLabels:
         app: online
     template:
       metadata:
         labels:
           app: online
       spec:
         schedulerName: volcano                 # Volcano is used.
         priorityClassName: volcano-production          # volcano-production priorityClass
         containers:
           - name: container-1
             image: resource_consumer:latest
             imagePullPolicy: IfNotPresent
             resources:
               requests:
                 cpu: 1400m
                 memory: 512Mi
               limits:
                 cpu: "2"
                 memory: 512Mi
         imagePullSecrets:
           - name: default-secret

   Online jobs are scheduled to the non-oversubscribed node.

   # kubectl get pod -o wide
   NAME                   READY   STATUS   RESTARTS  AGE     IP             NODE 
   online-ffb46f656-4mwr6  1/1    Running   0         5s    192.168.10.146   192.168.0.3
   online-ffb46f656-dqdv2   1/1    Running   0         5s    192.168.10.67   192.168.0.3

4. Improve the resource usage of the oversubscribed node and observe whether offline job eviction is triggered.
   Deploy online jobs to the oversubscribed node (192.168.0.173).

   apiVersion: apps/v1
   kind: Deployment
   metadata:
     name: online
     namespace: default
   spec:
     replicas: 2
     selector:
       matchLabels:
         app: online
     template:
       metadata:
         labels:
           app: online
       spec:
          affinity:                             # Submit an online job to an oversubscribed node.
           nodeAffinity:
             requiredDuringSchedulingIgnoredDuringExecution:
               nodeSelectorTerms:
               - matchExpressions:
                 - key: kubernetes.io/hostname
                   operator: In
                   values:
                   - 192.168.0.173
         schedulerName: volcano                 # Volcano is used.
         priorityClassName: volcano-production          # volcano-production priorityClass
         containers:
           - name: container-1
             image: resource_consumer:latest
             imagePullPolicy: IfNotPresent
             resources:
               requests:
                 cpu: 700m
                 memory: 512Mi
               limits:
                 cpu: 700m
                 memory: 512Mi
         imagePullSecrets:
           - name: default-secret

   Submit the online or offline jobs to the oversubscribed node (192.168.0.173) at the same time.

   # kubectl get pod -o wide
   NAME                     READY   STATUS   RESTARTS  AGE     IP             NODE 
   offline-69cdd49bf4-pmjp8  1/1     Running    0      13m   192.168.10.178   192.168.0.173 
   offline-69cdd49bf4-z8kxh  1/1     Running     0      13m   192.168.10.131   192.168.0.173 
   online-6f44bb68bd-b8z9p  1/1     Running     0     3m4s   192.168.10.18   192.168.0.173 
   online-6f44bb68bd-g6xk8  1/1     Running     0     3m12s   192.168.10.69   192.168.0.173

   Check the oversubscribed node with IP address 192.168.0.173. It is found that resources are oversubscribed, where there are 2343m CPU cores and 3073653200 MiB of memory. Additionally, the CPU allocation rate exceeded 100%.

   # kubectl describe node 192.168.0.173
   Name:              192.168.0.173
   Roles:              <none>
   Labels:              …
                       volcano.sh/oversubscription=true
   Annotations:         …                  
                       volcano.sh/oversubscription-cpu: 2343
                       volcano.sh/oversubscription-memory: 3073653200
                       …
   Allocated resources:
     (Total limits may be over 100 percent, i.e., overcommitted.)
     Resource               Requests      Limits
     --------               --------        ------
     cpu                    4750m (121%)  7350m (187%)
     memory                 3760Mi (61%)  4660Mi (76%)
                            …

   Increase the CPU usage of online jobs on the node. Offline job eviction is triggered.

   # kubectl get pod -o wide
   NAME                     READY   STATUS   RESTARTS  AGE     IP             NODE 
   offline-69cdd49bf4-bwdm7  1/1    Running   0       11m   192.168.10.208  192.168.0.3 
   offline-69cdd49bf4-pmjp8   0/1    Evicted    0       26m   <none>         192.168.0.173
   offline-69cdd49bf4-qpdss   1/1     Running   0       11m   192.168.10.174  192.168.0.3  
   offline-69cdd49bf4-z8kxh   0/1     Evicted    0       26m   <none>        192.168.0.173
   online-6f44bb68bd-b8z9p   1/1     Running   0       24m   192.168.10.18  192.168.0.173
   online-6f44bb68bd-g6xk8   1/1     Running   0       24m   192.168.10.69  192.168.0.173

Handling Suggestions

• After kubelet of the oversubscribed node is restarted, the resource view of Volcano Scheduler is not synchronized with that of kubelet. As a result, OutOfCPU occurs in some newly scheduled jobs, which is normal. After a period of time, Volcano Scheduler can properly schedule online and offline jobs.
• After online and offline jobs are submitted, you are not advised to dynamically change the job type (adding or deleting annotation volcano.sh/qos-level: "-1") because the current kernel does not support the change of an offline job to an online job.
• CCE collects the resource usage (CPU/memory) of all pods running on a node based on the status information in the cgroups system. The resource usage may be different from the monitored resource usage, for example, the resource statistics displayed by running the top command.
• You can add oversubscribed resources (such as CPU and memory) at any time.

  You can reduce the oversubscribed resource types only when the resource allocation rate does not exceed 100%.

• If an offline job is deployed on a node ahead of an online job and the online job cannot be scheduled due to insufficient resources, configure a higher priorityClass for the online job than that for the offline job.
• If there are only online jobs on a node and the eviction threshold is reached, the offline jobs that are scheduled to the current node will be evicted soon. This is normal.

Switching kubelet Oversubscription to Resource Oversubscription in Cloud Native Hybrid Deployment

If the cluster meets the migration requirements described in Table 1, perform the following operations to migrate oversubscription:

1. Enable resource oversubscription in cloud native hybrid deployment for a node pool that uses kubelet oversubscription. Either resource oversubscription in cloud native hybrid deployment or kubelet oversubscription can be used. Disable kubelet oversubscription for switching.
2. Disable both node oversubscription (over-subscription-resource) and hybrid deployment (colocation) of the node pool. Then, kubelet oversubscription will be automatically switched to resource oversubscription in cloud native hybrid deployment. During the switchover, kubelet temporarily cancels resource oversubscription reported through node annotations and evicts offline pods until the node's resource allocation rate is less than 100%. This is a normal case. After volcano-agent takes over resource oversubscription, this function will automatically restore.