Volcano Scheduler

Introduction

Volcano is a batch processing platform based on Kubernetes. It provides a series of features required by machine learning, deep learning, bioinformatics, genomics, and other big data applications, as a powerful supplement to Kubernetes capabilities.

Volcano provides general computing capabilities such as high-performance job scheduling, heterogeneous chip management, and job running management. It accesses the computing frameworks for various industries such as AI, big data, gene, and rendering and schedules up to 1000 pods per second for end users, greatly improving scheduling efficiency and resource utilization.

Volcano provides job scheduling, job management, and queue management for computing applications. Its main features are as follows:

• Diverse computing frameworks, such as TensorFlow, MPI, and Spark, can run on Kubernetes in containers. Common APIs for batch computing jobs through CRD, various plugins, and advanced job lifecycle management are provided.
• Advanced scheduling capabilities are provided for batch computing and high-performance computing scenarios, including group scheduling, preemptive priority scheduling, packing, resource reservation, and task topology.
• Queues can be effectively managed for scheduling jobs. Complex job scheduling capabilities such as queue priority and multi-level queues are supported.

Volcano has been open-sourced in GitHub at https://github.com/volcano-sh/volcano.

Install and configure the Volcano add-on in CCE clusters. For details, see Volcano Scheduling.

When using Volcano as a scheduler, use it to schedule all workloads in the cluster. This prevents resource scheduling conflicts caused by simultaneous working of multiple schedulers.

Notes and Constraints

If the Volcano Scheduler add-on is upgraded from 1.4.7 or earlier to a version later than 1.4.7, the webhooks.admissionReviewVersions field information in the new version may be incompatible with that in the old version. As a result, VolcanoJob (vcjob) resources cannot be created.

Installing the Add-on

1. Log in to the CCE console and click the cluster name to access the cluster console.
2. In the navigation pane, choose Add-ons. Locate Volcano Scheduler on the right and click Install.
3. On the Install Add-on page, configure the specifications as needed.
   • If you selected Preset, the system will configure the number of pods and resource quotas for the add-on based on the preset specifications. You can see the configurations on the console.
   • If you selected Custom, you can adjust the number of pods and resource quotas as needed. High availability is not possible with a single pod. If an error occurs on the node where the add-on instance runs, the add-on will fail.

     The resource quotas of the volcano-admission component are related to the cluster scale. For details, see Table 1. The resource quotas of volcano-controller and volcano-scheduler are related to the number of cluster nodes and pods. The recommended values are as follows:

     • If the number of nodes is less than 100, retain the default configuration. The requested vCPUs are 500m, and the limit is 2000m. The requested memory is 500 MiB, and the limit is 2000 MiB.
     • If the number of nodes is greater than 100, increase the requested vCPUs by 500m and the requested memory by 1000 MiB each time 100 nodes (10,000 pods) are added. Increase the vCPU limit by 1500m and the memory limit by 1000 MiB.
       

       Recommended formula for calculating the requested value:

       • Requested vCPUs: Calculate the number of target nodes multiplied by the number of target pods, perform interpolation search based on the number of nodes in the cluster multiplied by the number of target pods in Table 2, and round up the request value and limit value that are closest to the specifications.

         For example, for 2000 nodes and 20,000 pods, Number of target nodes x Number of target pods = 40 million, which is close to the specification of 700/70,000 (Number of cluster nodes x Number of pods = 49 million). According to the following table, set the requested vCPUs to 4000m and the limit value to 5500m.

       • Requested memory: It is recommended that 2.4 GiB memory be allocated to every 1000 nodes and 1 GiB memory be allocated to every 10,000 pods. The requested memory is the sum of these two values. (The obtained value may be different from the recommended value in Table 2. You can use either of them.)

         Requested memory = Number of target nodes/1000 × 2.4 GiB + Number of target pods/10,000 × 1 GiB

         For example, for 2000 nodes and 20,000 pods, the requested memory is 6.8 GiB (2000/1000 × 2.4 GiB + 20,000/10,000 × 1 GiB).

     Table 1 Recommended requested resources and resource limits for volcano-admission

     Cluster Scale	CPU Request (m)	vCPU Limit (m)	Memory Request (MiB)	Memory Limit (MiB)
     50 nodes	200	500	500	500
     200 nodes	500	1000	1000	2000
     1000 or more nodes	1500	2500	3000	4000

     Table 2 Recommended requested resources and resource limits for volcano-controller and volcano-scheduler

     Nodes/Pods in a Cluster	CPU Request (m)	CPU Limit (m)	Memory Request (MiB)	Memory Limit (MiB)
     50/5000	500	2000	500	2000
     100/10000	1000	2500	1500	2500
     200/20000	1500	3000	2500	3500
     300/30000	2000	3500	3500	4500
     400/40000	2500	4000	4500	5500
     500/50000	3000	4500	5500	6500
     600/60000	3500	5000	6500	7500
     700/70000	4000	5500	7500	8500

4. Configure the extended functions supported by the add-on.
   • Descheduling: After this function is enabled, the volcano-descheduler component is automatically deployed. The scheduler will evict and reschedule pods that do not meet your policy configuration requirements. This helps to balance cluster load and reduce resource fragmentation. For details, see Descheduling.
   • NUMA Topology Scheduling: After this function is enabled, the resource-exporter component is automatically deployed. The scheduler will schedule pods in NUMA affinity mode, which enhances the performance of high-performance training jobs. For details, see NUMA Affinity Scheduling.

5. Configure deployment policies for the add-on pods.
   

   • Scheduling policies do not take effect on add-on pods of the DaemonSet type.
   • When configuring multi-AZ deployment or node affinity, ensure that there are nodes meeting the scheduling policy and that resources are sufficient in the cluster. Otherwise, the add-on cannot run.

   Table 3 Configurations for add-on scheduling

   Parameter	Description
   Multi-AZ Deployment	Preferred: Deployment pods of the add-on will be preferentially scheduled to nodes in different AZs. If all the nodes in the cluster are deployed in the same AZ, the pods will be scheduled to different nodes in that AZ. Forcible: Deployment pods of the add-on are forcibly scheduled to nodes in different AZs. There can be at most one pod in each AZ. If nodes in a cluster are not in different AZs, some add-on pods cannot run properly. If a node is faulty, add-on pods on it may fail to be migrated.
   Node Affinity	Not configured: Node affinity is disabled for the add-on. Specify node: Specify the nodes where the add-on is deployed. If you do not specify the nodes, the add-on will be randomly scheduled based on the default cluster scheduling policy. Specify node pool: Specify the node pool where the add-on is deployed. If you do not specify the node pools, the add-on will be randomly scheduled based on the default cluster scheduling policy. Customize affinity: Enter the labels of the nodes where the add-on is to be deployed for more flexible scheduling policies. If you do not specify node labels, the add-on will be randomly scheduled based on the default cluster scheduling policy. If multiple custom affinity policies are configured, ensure that there are nodes that meet all the affinity policies in the cluster. Otherwise, the add-on cannot run.
   Toleration	Using both taints and tolerations allows (not forcibly) the add-on Deployment to be scheduled to a node with the matching taints, and controls the Deployment eviction policies after the node where the Deployment is located is tainted. The add-on adds the default tolerance policy for the node.kubernetes.io/not-ready and node.kubernetes.io/unreachable taints, respectively. The tolerance time window is 60s. For details, see Configuring Tolerance Policies.

6. Click Install.
   After the add-on is installed, you can choose Settings in the navigation pane, switch to the Scheduling tab, and find the expert mode. You can customize advanced scheduling policies based on actual service scenarios. The following is an example:

   admission_kube_api_qps: 200
   admissions: /jobs/mutate,/jobs/validate,/podgroups/mutate,/pods/validate,/pods/mutate,/queues/mutate,/queues/validate,/eas/pods/mutate,/eas/pods/validate,/npu/jobs/validate,/resource/validate,/resource/mutate,/workloadbalancer/balancer/validate,/workloadbalancer/balancerpolicytemplate/validate
   annotations: {}
   colocation_enable: 'false'
   controller_kube_api_qps: 200
   default_scheduler_conf:
     actions: allocate, backfill, preempt
     metrics:
       interval: 30s
       type: ''
     tiers:
       - plugins:
           - name: priority
           - enableJobStarving: false
             enablePreemptable: false
             name: gang
           - name: conformance
       - plugins:
           - enablePreemptable: false
             name: drf
           - name: predicates
           - name: nodeorder
       - plugins:
           - name: cce-gpu-topology-predicate
           - name: cce-gpu-topology-priority
           - name: xgpu
       - plugins:
           - name: nodelocalvolume
           - name: nodeemptydirvolume
           - name: nodeCSIscheduling
           - name: networkresource
   deschedulerPolicy:
     profiles:
       - name: ProfileName
         pluginConfig:
           - args:
               nodeFit: true
             name: DefaultEvictor
           - args:
               evictableNamespaces:
                 exclude:
                   - kube-system
               thresholds:
                 cpu: 20
                 memory: 20
             name: HighNodeUtilization
           - args:
               evictableNamespaces:
                 exclude:
                   - kube-system
               metrics:
                 type: prometheus_adaptor
               nodeFit: true
               targetThresholds:
                 cpu: 80
                 memory: 85
               thresholds:
                 cpu: 30
                 memory: 30
             name: LoadAware
         plugins:
           balance:
             enabled: null
   descheduler_enable: 'false'
   deschedulingInterval: 10m
   enable_workload_balancer: false
   oversubscription_method: nodeResource
   oversubscription_profile_period: 300
   oversubscription_ratio: 60
   recommendation_enable: ''
   scheduler_kube_api_qps: 200
   update_pod_status_qps: 50
   workload_balancer_score_annotation_key: ''
   workload_balancer_third_party_types: ''

   Table 4 Advanced Volcano configuration parameters

   Function	Parameter	Function	Description
   Basic scheduling functions	admission_kube_api_qps	QPS of requests sent by volcano-admission to Kubernetes API server	Default value: 200; parameter type: float
   	controller_kube_api_qps	QPS of requests sent by volcano-controller to Kubernetes API server	Default value: 200; parameter type: float
   	scheduler_kube_api_qps	QPS of requests sent by volcano-scheduler to Kubernetes API server	Default value: 200; parameter type: float
   	update_pod_status_qps	QPS of the requests for updating the pod status by volcano-scheduler	Default value: 50; parameter type: float
   	default_scheduler_conf	Used to schedule pods. It consists of a series of actions and plugins and features high scalability. You can specify and implement actions and plugins based on your requirements.	It consists of: actions: defines the types and sequence of actions to be executed by the scheduler. tiers: configures the plugin list.
   	default_scheduler_conf.actions	Actions to be executed in each scheduling phase. The configured action sequence is the scheduler execution sequence. For details, see Actions. The scheduler traverses all jobs to be scheduled and performs actions such as enqueue, allocate, preempt, and backfill in the configured sequence to find the most appropriate node for each job.	The following options are supported: enqueue: uses a series of filtering algorithms to filter out tasks to be scheduled and sends them to the queue to wait for scheduling. After this action, the task status changes from pending to inqueue. allocate: selects the most suitable node based on a series of pre-selection and selection algorithms. preempt: performs preemption scheduling for tasks with higher priorities in the same queue based on priority rules. backfill: schedules tasks in the pending state as much as possible to maximize node resource utilization. Example: actions: 'allocate, backfill, preempt' NOTE: When configuring actions, use either preempt or enqueue.
   	default_scheduler_conf.tier.plugin	Implementation details of algorithms in actions based on different scenarios. For details, see Plugins.	For details, see Table 5.
   Descheduling	descheduler_enable	Used to enable descheduling.	This function is disabled by default. Options: true: The function is enabled. false or empty: The function is disabled.
   	deschedulerPolicy	Descheduling policy	For details about the parameters, see Table 2.
   	deschedulingInterval	Descheduling period	Value range: > 0s; parameter type: time
   Cloud native hybrid deployment	colocation_enable	Used to enable cloud native hybrid deployment.	This function is disabled by default. Options: true: The function is enabled. false or empty: The function is disabled.
   	oversubscription_method	Method for calculating the oversubscription	nodeResource and podProfile are supported. The default value is nodeResource. nodeResource: calculates the oversubscription based on the node resource usage. podProfile: calculates the oversubscription based on pod profiling.
   	oversubscription_ratio	Percentage of idle resource oversubscription of a node	Value range: 1 to 100; parameter type: int For example, 60 indicates that the maximum oversubscription resources on a node is calculated based on 60% × Idle resources on the node.
   	oversubscription_profile_period	Period of pod profiling	Value range: 60 to 2592000, in seconds, that is, from 1 minute to 1 month. If a pod's metrics are not collected for the entire period, the node's resources will be evaluated according to the resources requested by the pod. When the oversubscription algorithm based on pod profiling is enabled for the first time, the amount of collected data may not be sufficient to cover the entire period. In this case, the oversubscription on the node is temporarily 0 due to lack of initialization data. After the data of the first period is collected, the oversubscription is updated to the actual value.

   Table 5 Supported plugins

   Plugins	Function	Description	Demonstration
   binpack	Schedule pods to nodes with high resource usage (not allocating pods to light-loaded nodes) to reduce resource fragments.	arguments: binpack.weight: weight of the binpack plugin. binpack.cpu: ratio of CPUs to all resources. The parameter value defaults to 1. binpack.memory: ratio of memory resources to all resources. The parameter value defaults to 1. binpack.resources: other custom resource types requested by the pod, for example, nvidia.com/gpu. Multiple types can be configured and be separated by commas (,). binpack.resources.<your_resource>: weight of your custom resource in all resources. Multiple types of resources can be added. <your_resource> indicates the resource type defined in binpack.resources, for example, binpack.resources.nvidia.com/gpu.	- plugins: - name: binpack arguments: binpack.weight: 10 binpack.cpu: 1 binpack.memory: 1 binpack.resources: nvidia.com/gpu, example.com/foo binpack.resources.nvidia.com/gpu: 2 binpack.resources.example.com/foo: 3
   conformance	Prevent key pods, such as the pods in the kube-system namespace from being preempted.	None	- plugins: - name: 'priority' - name: 'gang' enablePreemptable: false - name: 'conformance'
   lifecycle	By collecting statistics on service scaling rules, pods with similar lifecycles are preferentially scheduled to the same node. With the horizontal scaling capability of the Autoscaler, resources can be quickly scaled in and released, reducing costs and improving resource utilization. 1. Collects statistics on the lifecycle of pods in the service load and schedules pods with similar lifecycles to the same node. 2. For a cluster configured with an automatic scaling policy, adjust the scale-in annotation of the node to preferentially scale in the node with low usage.	arguments: lifecycle.WindowSize: The value is an integer greater than or equal to 1 and defaults to 10. Record the number of times that the number of replicas changes. If the load changes regularly and periodically, decrease the value. If the load changes irregularly and the number of replicas changes frequently, increase the value. If the value is too large, the learning period is prolonged and too many events are recorded. lifecycle.MaxGrade: The value is an integer greater than or equal to 3 and defaults to 3. It indicates levels of replicas. For example, if the value is set to 3, the replicas are classified into three levels. If the load changes regularly and periodically, decrease the value. If the load changes irregularly, increase the value. Setting an excessively small value may result in inaccurate lifecycle forecasts. lifecycle.MaxScore: float64 floating point number. The value must be greater than or equal to 50.0. The default value is 200.0. Maximum score (equivalent to the weight) of the lifecycle plugin. lifecycle.SaturatedTresh: float64 floating point number. If the value is less than 0.5, use 0.5. If the value is greater than 1, use 1. The default value is 0.8. Threshold for determining whether the node usage is too high. If the node usage exceeds the threshold, the scheduler preferentially schedules jobs to other nodes.	- plugins: - name: priority - name: gang enablePreemptable: false - name: conformance - name: lifecycle arguments: lifecycle.MaxGrade: 3 lifecycle.MaxScore: 200.0 lifecycle.SaturatedTresh: 0.8 lifecycle.WindowSize: 10 NOTE: For nodes that do not want to be scaled in, manually mark them as long-period nodes and add the annotation volcano.sh/long-lifecycle-node: true to them. For an unmarked node, the lifecycle plugin automatically marks the node based on the lifecycle of the load on the node. The default value of MaxScore is 200.0, which is twice the weight of other plugins. When the lifecycle plugin does not have obvious effect or conflicts with other plugins, disable other plugins or increase the value of MaxScore. After the scheduler is restarted, the lifecycle plugin needs to re-record the load change. The optimal scheduling effect can be achieved only after several periods of statistics are collected.
   Gang	Consider a group of pods as a whole for resource allocation. This plugin checks whether the number of scheduled pods in a job meets the minimum requirements for running the job. If yes, all pods in the job will be scheduled. If no, the pods will not be scheduled. NOTE: If a gang scheduling policy is used, if the remaining resources in the cluster are greater than or equal to half of the minimum number of resources for running a job but less than the minimum of resources for running the job, Autoscaler scale-outs will not be triggered.	enablePreemptable: true: Preemption enabled false: Preemption not enabled enableJobStarving: true: Resources are preempted based on the minAvailable setting of jobs. false: Resources are preempted based on job replicas. NOTE: The default value of minAvailable for Kubernetes-native workloads (such as Deployments) is 1. It is a good practice to set enableJobStarving to false. In AI and big data scenarios, you can specify the minAvailable value when creating a vcjob. It is a good practice to set enableJobStarving to true. In Volcano versions earlier than v1.11.5, enableJobStarving is set to true by default. In Volcano versions later than v1.11.5, enableJobStarving is set to false by default.	- plugins: - name: priority - name: gang enablePreemptable: false enableJobStarving: false - name: conformance
   priority	Schedule based on custom load priorities.	None	- plugins: - name: priority - name: gang enablePreemptable: false - name: conformance
   overcommit	Resources in a cluster are scheduled after being accumulated in a certain multiple to improve the workload enqueuing efficiency. If all workloads are Deployments, remove this plugin or set the raising factor to 2.0. NOTE: This plugin is supported in Volcano 1.6.5 and later versions.	arguments: overcommit-factor: inflation factor, which defaults to 1.2.	- plugins: - name: overcommit arguments: overcommit-factor: 2.0
   drf	The Dominant Resource Fairness (DRF) scheduling algorithm, which schedules jobs based on their dominant resource share. Jobs with a smaller resource share will be scheduled with a higher priority.	None	- plugins: - name: 'drf' - name: 'predicates' - name: 'nodeorder'
   predicates	Determine whether a task is bound to a node by using a series of evaluation algorithms, such as node/pod affinity, taint tolerance, node repetition, volume limits, and volume zone matching.	None	- plugins: - name: 'drf' - name: 'predicates' - name: 'nodeorder'
   nodeorder	A common algorithm for selecting nodes. Nodes are scored in simulated resource allocation to find the most suitable node for the current job.	Scoring parameters: nodeaffinity.weight: Pods are scheduled based on node affinity. This parameter defaults to 2. podaffinity.weight: Pods are scheduled based on pod affinity. This parameter defaults to 2. leastrequested.weight: Pods are scheduled to the node with the least requested resources. This parameter defaults to 1. balancedresource.weight: Pods are scheduled to the node with balanced resource allocation. This parameter defaults to 1. mostrequested.weight: Pods are scheduled to the node with the most requested resources. This parameter defaults to 0. tainttoleration.weight: Pods are scheduled to the node with a high taint tolerance. This parameter defaults to 3. imagelocality.weight: Pods are scheduled to the node where the required images exist. This parameter defaults to 1. podtopologyspread.weight: Pods are scheduled based on the pod topology. This parameter defaults to 2.	- plugins: - name: nodeorder arguments: leastrequested.weight: 1 mostrequested.weight: 0 nodeaffinity.weight: 2 podaffinity.weight: 2 balancedresource.weight: 1 tainttoleration.weight: 3 imagelocality.weight: 1 podtopologyspread.weight: 2
   cce-gpu-topology-predicate	GPU-topology scheduling preselection algorithm	None	- plugins: - name: 'cce-gpu-topology-predicate' - name: 'cce-gpu-topology-priority' - name: 'xgpu'
   cce-gpu-topology-priority	GPU-topology scheduling priority algorithm	None	- plugins: - name: 'cce-gpu-topology-predicate' - name: 'cce-gpu-topology-priority' - name: 'xgpu'
   cce-gpu	GPU resource allocation that supports decimal GPU configurations by working with the CCE AI Suite (NVIDIA GPU) add-on.	None	- plugins: - name: 'cce-gpu-topology-predicate' - name: 'cce-gpu-topology-priority' - name: 'cce-gpu'
   numa-aware	NUMA affinity scheduling. For details, see NUMA Affinity Scheduling.	arguments: weight: weight of the numa-aware plugin	- plugins: - name: 'nodelocalvolume' - name: 'nodeemptydirvolume' - name: 'nodeCSIscheduling' - name: 'networkresource' arguments: NetworkType: vpc-router - name: numa-aware arguments: weight: 10
   networkresource	Filter out nodes that require elastic network interfaces. The parameters are transferred by CCE and do not need to be manually configured.	arguments: NetworkType: network type (eni or vpc-router)	- plugins: - name: 'nodelocalvolume' - name: 'nodeemptydirvolume' - name: 'nodeCSIscheduling' - name: 'networkresource' arguments: NetworkType: vpc-router
   nodelocalvolume	Filter out nodes that do not meet local volume requirements.	None	- plugins: - name: 'nodelocalvolume' - name: 'nodeemptydirvolume' - name: 'nodeCSIscheduling' - name: 'networkresource'
   nodeemptydirvolume	Filter out nodes that do not meet the emptyDir requirements.	None	- plugins: - name: 'nodelocalvolume' - name: 'nodeemptydirvolume' - name: 'nodeCSIscheduling' - name: 'networkresource'
   nodeCSIscheduling	Filter out nodes with malfunctional Everest.	None	- plugins: - name: 'nodelocalvolume' - name: 'nodeemptydirvolume' - name: 'nodeCSIscheduling' - name: 'networkresource'

Components

Modifying the volcano-scheduler Configurations Using the Console

volcano-scheduler is the component responsible for pod scheduling. It consists of a series of actions and plugins. Actions should be executed in every step. Plugins provide the action algorithm details in different scenarios. volcano-scheduler is highly scalable. You can specify and implement actions and plugins based on your requirements.

After the add-on is installed, you can choose Settings in the navigation pane, switch to the Scheduling tab, and configure the basic scheduling capabilities. You can also use the expert mode to customize advanced scheduling policies based on service scenarios.

This section describes how to configure volcano-scheduler.

Only Volcano of v1.7.1 and later support this function.

Log in to the CCE console and click the cluster name to access the cluster console. In the navigation pane, choose Settings and click the Scheduling tab. In the Default Cluster Scheduler area, find the expert mode and click Try Now.

• Using resource_exporter:

  ...
      "default_scheduler_conf": {
          "actions": "allocate, backfill, preempt",
          "tiers": [
              {
                  "plugins": [
                      {
                          "name": "priority"
                      },
                      {
                          "name": "gang"
                      },
                      {
                          "name": "conformance"
                      }
                  ]
              },
              {
                  "plugins": [
                      {
                          "name": "drf"
                      },
                      {
                          "name": "predicates"
                      },
                      {
                          "name": "nodeorder"
                      }
                  ]
              },
              {
                  "plugins": [
                      {
                          "name": "cce-gpu-topology-predicate"
                      },
                      {
                          "name": "cce-gpu-topology-priority"
                      },
                      {
                          "name": "cce-gpu"
                      },
                      {
                          "name": "numa-aware" # add this also enable resource_exporter
                      }
                  ]
              },
              {
                  "plugins": [
                      {
                          "name": "nodelocalvolume"
                      },
                      {
                          "name": "nodeemptydirvolume"
                      },
                      {
                          "name": "nodeCSIscheduling"
                      },
                      {
                          "name": "networkresource"
                      }
                  ]
              }
          ]
      },
  ...

  After this function is enabled, you can use the functions of both numa-aware and resource_exporter.

Collecting Prometheus Metrics

volcano-scheduler exposes Prometheus metrics through port 8080. You can build a Prometheus collector to identify and obtain volcano-scheduler scheduling metrics from http://{{volcano-schedulerPodIP}}:{{volcano-schedulerPodPort}}/metrics.

Prometheus metrics can be exposed only by the Volcano add-on of version 1.8.5 or later.

Uninstalling the Volcano Add-on

After the add-on is uninstalled, all custom Volcano resources (Table 8) will be deleted, including the created resources. Reinstalling the add-on will not inherit or restore the tasks before the uninstallation. It is a good practice to uninstall the Volcano add-on only when no custom Volcano resources are being used in the cluster.

Release History

It is a good practice to upgrade Volcano to the latest version that is supported by the cluster.