Dynamic Resource Oversubscription¶
Many services see surges in traffic. To ensure performance and stability, resources are often requested at the maximum needed. However, the surges may ebb very shortly and resources, if not released, are wasted in non-peak hours. Especially for online jobs that request a large quantity of resources to ensure SLA, resource utilization can be as low as it gets.
Resource oversubscription is the process of making use of idle requested resources. Oversubscribed resources are suitable for deploying offline jobs, which focus on throughput but have low SLA requirements and can tolerate certain failures.
Hybrid deployment of online and offline jobs in a cluster can better utilize cluster resources.
Figure 1 Resource oversubscription¶
Features¶
Note
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.
Compatible kubelet Oversubscription¶
Important
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 HCE OS 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 1.
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.
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. For details about the related configuration, see Table 1.
Ensure that you have correctly configure labels because the scheduler does not check the add-on and node configurations.
Oversubscription in Add-on | Oversubscription Label on Node | Scheduling |
|---|---|---|
Yes | Yes | Triggered by oversubscription |
Yes | No | Triggered |
No | No | Triggered |
No | Yes | Not triggered or failed. Avoid this configuration. |
Use kubectl to access the cluster.
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 ...
Enable node oversubscription.
A label can be configured to use oversubscribed resources only after the oversubscription feature is enabled for a node. Related nodes can be created only in a node pool. To enable the oversubscription feature, perform the following steps:
Create a node pool.
Choose Manage in the Operation column of the created node pool.
On the Manage Configurations page, enable Node oversubscription feature (over-subscription-resource) and click OK.
Set the node oversubscription label.
The volcano.sh/oversubscription label needs to be configured for an oversubscribed node. If this label is set for a node and the value is true, the node is an oversubscribed node. Otherwise, the node is not an oversubscribed node.
kubectl label node 192.168.0.0 volcano.sh/oversubscription=true
An oversubscribed node also supports the oversubscription thresholds, as listed in Table 2. For example:
kubectl annotate node 192.168.0.0 volcano.sh/evicting-cpu-high-watermark=70
Querying the node information
# kubectl describe node 192.168.0.0 Name: 192.168.0.0 Roles: <none> Labels: ... volcano.sh/oversubscription=true Annotations: ... volcano.sh/evicting-cpu-high-watermark: 70
Table 2 Node oversubscription annotations¶ Parameter
Description
volcano.sh/evicting-cpu-high-watermark
Upper limit for CPU usage. 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%.
volcano.sh/evicting-cpu-low-watermark
Lower limit for CPU usage. 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%.
volcano.sh/evicting-memory-high-watermark
Upper limit for memory usage. 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%.
volcano.sh/evicting-memory-low-watermark
Lower limit for memory usage. 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%.
volcano.sh/oversubscription-types
Oversubscribed resource type. Options:
cpu: oversubscribed CPU
memory: oversubscribed memory
cpu,memory: oversubscribed CPU and memory
The default value is cpu,memory.
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
Deploy online and offline jobs and configure priorityClasses for these jobs.
The volcano.sh/qos-level annotation needs to be added to distinguish offline jobs. The value is an integer ranging from -7 to 7. If the value is less than 0, the job is an offline job. If the value is greater than or equal to 0, the job is an online job. You do not need to set this annotation for online jobs. For both online and offline jobs, set schedulerName to volcano to enable Volcano.
Note
The priorities between online jobs and between offline jobs are not differentiated, and the value validity is not verified. If the value of volcano.sh/qos-level of an offline job is not a negative integer ranging from -7 to 0, the job is processed as an online job.
For an offline job:
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 an online job:
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 ...
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%)
In the preceding command, CPU and memory are in the unit of m CPU cores and MiB, respectively.
Deployment Example¶
The following uses an example to describe how to deploy online and offline jobs in hybrid mode.
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
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
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
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.