Descheduling

Scheduling in a cluster is the process of binding pending pods to nodes, and is performed by a component called kube-scheduler or Volcano Scheduler. The scheduler uses a series of algorithms to compute the optimal node for running pods. However, Kubernetes clusters are dynamic and their state changes over time. For example, if a node needs to be maintained, all pods on the node will be evicted to other nodes. After the maintenance is complete, the evicted pods will not automatically return back to the node because descheduling will not be triggered once a pod is bound to a node. Due to these changes, the load of a cluster may be unbalanced after the cluster runs for a period of time.

CCE has resolved this issue by using Volcano Scheduler to evict pods that do not comply with the configured policy so that pods can be rescheduled. In this way, the cluster load is balanced and resource fragmentation is minimized.

Features

Load-aware Descheduling

During Kubernetes cluster management, over-utilized nodes are due to high CPU or memory usage, which affects the stable running of pods on these nodes and increases the probability of node faults. To dynamically balance the resource usage between nodes in a cluster, a cluster resource view is required based on node monitoring metrics. During cluster management, real-time monitoring can be used to detect issues such as high resource usage on a node, node faults, and excessive number of pods on a node so that the system can take measures promptly, for example, by migrating some pods from an over-utilized node to under-utilized nodes.

Figure 1 Load-aware descheduling

When using this add-on, ensure the highThresholds value is greater than the lowThresholds value. Otherwise, the descheduler cannot work.

Appropriately utilized node: a node whose resource usage is greater than or equal to 30% and less than or equal to 80%. The resource usage of appropriately utilized nodes is within the expected range.
Over-utilized node: a node whose resource usage is higher than 80%. Some pods will be evicted from over-utilized nodes to reduce its resource usage to be less than or equal to 80%. The descheduler will schedule the evicted pods to under-utilized nodes.
Under-utilized node: a node whose resource usage is lower than 30%.

HighNodeUtilization

This policy finds nodes that are under-utilized and evicts pods from the nodes in the hope that these pods will be scheduled compactly into fewer nodes. This policy must be used with the bin packing policy of Volcano Scheduler or the MostAllocated policy of the kube-scheduler scheduler. Thresholds can be configured for CPU and memory.

Prerequisites

A cluster of v1.19.16 or later is available. For details, see Creating a CCE Standard Cluster.
Volcano of v1.11.5 or later has been installed in the cluster. For details, see Volcano Scheduler.

Constraints

Pods need to be rescheduled using a scheduler, and no scheduler can label pods or nodes. Therefore, an evicted pod might be rescheduled to the original node.
Descheduling does not support anti-affinity between pods. An evicted pod is in anti-affinity relationship with other running pods. Therefore, the scheduler may still schedule the pod back to the node where the pod was evicted from.
When configuring load-aware descheduling, you are required to enable load-aware scheduling on Volcano Scheduler. When configuring HighNodeUtilization, you are required to enable bin packing on Volcano Scheduler.

Configuring a Load-aware Descheduling Policy

When configuring a load-aware descheduling policy, do as follows to enable load-aware scheduling on Volcano Scheduler:

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 on the right side of the page. Then, enable load-aware scheduling.
In the navigation pane, choose Add-ons. Locate Volcano Scheduler on the right and click Install or Edit.

In the Parameters area, modify Advanced Settings to configure the load-aware descheduling policy. The following shows a configuration example for Volcano 1.11.21 or later:

{
  "colocation_enable": "",
  "default_scheduler_conf": {
    "actions": "allocate, backfill, preempt",
    "tiers": [
      {
        "plugins": [
          {
            "name": "priority"
          },
          {
            "enablePreemptable": false,
            "name": "gang"
          },
          {
            "name": "conformance"
          }
        ]
      },
      {
        "plugins": [
          {
            "enablePreemptable": false,
            "name": "drf"
          },
          {
            "name": "predicates"
          },
          {
            "name": "nodeorder"
          },
          {
            "name": "usage",
            "enablePredicate": true,
            "arguments": {
              "usage.weight": 5,
              "cpu.weight": 1,
              "memory.weight": 1,
              "thresholds": {
                "cpu": 80,
                "mem": 80
              }
            }
          }
        ]
      },
      {
        "plugins": [
          {
            "name": "cce-gpu-topology-predicate"
          },
          {
            "name": "cce-gpu-topology-priority"
          },
          {
            "name": "cce-gpu"
          }
        ]
      },
      {
        "plugins": [
          {
            "name": "nodelocalvolume"
          },
          {
            "name": "nodeemptydirvolume"
          },
          {
            "name": "nodeCSIscheduling"
          },
          {
            "name": "networkresource"
          }
        ]
      }
    ]
  },
  "deschedulerPolicy": {
    "profiles": [
      {
        "name": "ProfileName",
        "pluginConfig": [
          {
            "args": {
              "ignorePvcPods": true,
              "nodeFit": true,
              "priorityThreshold": {
                "value": 100
              }
            },
            "name": "DefaultEvictor"
          },
          {
            "args": {
              "evictableNamespaces": {
                "exclude": ["kube-system"]
              },
              "metrics": {
                "type": "prometheus_adaptor"
              },
              "targetThresholds": {
                "cpu": 80,
                "memory": 85
              },
              "thresholds": {
                "cpu": 30,
                "memory": 30
              }
            },
            "name": "LoadAware"
          }
        ],
        "plugins": {
          "balance": {
            "enabled": ["LoadAware"]
          }
        }
      }
    ]
  },
  "descheduler_enable": "true",
  "deschedulingInterval": "10m"
}

**Table 1** Key parameters of a cluster descheduling policy
Parameter	Description
descheduler_enable	Whether to enable a cluster descheduling policy. true: The cluster descheduling policy is enabled. false: The cluster descheduling policy is disabled.
deschedulingInterval	Descheduling period.
deschedulerPolicy	Cluster descheduling policy. For details, see Table 2.

**Table 2** deschedulerPolicy parameters
Parameter	Description
profiles.[].plugins.balance.enable.[]	Descheduling policy for a cluster. LoadAware: a load-aware descheduling policy is used.
profiles.[].pluginConfig.[].name	Configuration of a load-aware descheduling policy. Options: DefaultEvictor: default eviction policy LoadAware: a load-aware descheduling policy
profiles.[].pluginConfig.[].args	Descheduling policy configuration of a cluster. Configurations for the DefaultEvictor policy: ignorePvcPods: whether PVC pods should be ignored or evicted. Value true indicates that the pods are ignored, and value false indicates that the pods are evicted. This configuration does not differentiate PVC types (local PVs or EVS). nodeFit: whether to consider the existing scheduling configurations such as node affinity and taint on the node during descheduling. Value true indicates that the existing scheduling configurations will be considered, and value false indicates that those will be ignored. priorityThreshold: priority setting. If the priority of a pod is greater than or equal to the value of this parameter, the pod will not be evicted. Example: { "value": 100 } Configurations for the LoadAware policy: evictableNamespaces: namespaces where the eviction policy takes effect. The default value is the namespaces other than kube-system. Example: { "exclude": ["kube-system"] } metrics: how monitoring data is obtained. Either the Custom Metrics API (prometheus_adaptor) or Prometheus can be used. For Volcano 1.11.17 and later versions, use Custom Metrics API to obtain monitoring data. The following is an example: { "type": "prometheus_adaptor" } For Volcano 1.11.5 to 1.11.16, use Prometheus to obtain monitoring data. You need to enter the IP address of the Prometheus server. The following is an example: { "address": "http://10.247.119.103:9090", "type": "prometheus" } targetThresholds: threshold for evicting pods from a node. When the CPU or memory usage of a node is greater than the threshold, the pods on the node will be evicted. Example: { "cpu": 60, "memory": 65 } thresholds: threshold for a node to run pods. If the node value is less than the threshold, the node allows evicted pods to run. Example: { "cpu": 30, "memory": 30 }

Click OK.

Configuring a HighNodeUtilization Policy

When configuring a HighNodeUtilization policy, do as follows to enable the bin packing policy on Volcano Scheduler:

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 on the right side of the page. Then, enable bin packing. For details, see Bin Packing.
In the navigation pane, choose Add-ons. Locate Volcano Scheduler on the right and click Install or Edit.

In the Parameters area, modify Advanced Settings to configure the HighNodeUtilization policy.

{
  "colocation_enable": "",
  "default_scheduler_conf": {
    "actions": "allocate, backfill, preempt",
    "tiers": [
      {
        "plugins": [
          {
            "name": "priority"
          },
          {
            "enablePreemptable": false,
            "name": "gang"
          },
          {
            "name": "conformance"
          },
          {
            "arguments": {
              "binpack.weight": 5
            },
            "name": "binpack"
          }
        ]
      },
      {
        "plugins": [
          {
            "enablePreemptable": false,
            "name": "drf"
          },
          {
            "name": "predicates"
          },
          {
            "name": "nodeorder"
          }
        ]
      },
      {
        "plugins": [
          {
            "name": "cce-gpu-topology-predicate"
          },
          {
            "name": "cce-gpu-topology-priority"
          },
          {
            "name": "cce-gpu"
          }
        ]
      },
      {
        "plugins": [
          {
            "name": "nodelocalvolume"
          },
          {
            "name": "nodeemptydirvolume"
          },
          {
            "name": "nodeCSIscheduling"
          },
          {
            "name": "networkresource"
          }
        ]
      }
    ]
  },
  "deschedulerPolicy": {
    "profiles": [
      {
        "name": "ProfileName",
        "pluginConfig": [
          {
            "args": {
              "ignorePvcPods": true,
              "nodeFit": true,
              "priorityThreshold": {
                "value": 100
              }
            },
            "name": "DefaultEvictor"
          },
          {
            "args": {
              "evictableNamespaces": {
                "exclude": ["kube-system"]
              },
              "thresholds": {
                "cpu": 25,
                "memory": 25
              }
            },
            "name": "HighNodeUtilization"
          }
        ],
        "plugins": {
          "balance": {
            "enabled": ["HighNodeUtilization"]
          }
        }
      }
    ]
  },
  "descheduler_enable": "true",
  "deschedulingInterval": "10m"
}

**Table 3** Key parameters of a cluster descheduling policy
Parameter	Description
descheduler_enable	Whether to enable a cluster descheduling policy. true: The cluster descheduling policy is enabled. false: The cluster descheduling policy is disabled.
deschedulingInterval	Descheduling period.
deschedulerPolicy	Cluster descheduling policy. For details, see Table 4.

**Table 4** deschedulerPolicy parameters
Parameter	Description
profiles.[].plugins.balance.enable.[]	Descheduling policy for a cluster. HighNodeUtilization: the policy for minimizing CPU and memory fragments is used.
profiles.[].pluginConfig.[].name	Configuration of a load-aware descheduling policy. Options: DefaultEvictor: default eviction policy HighNodeUtilization: policy for minimizing CPU and memory fragments
profiles.[].pluginConfig.[].args	Descheduling policy configuration of a cluster. Configurations for the DefaultEvictor policy: ignorePvcPods: whether PVC pods should be ignored or evicted. Value true indicates that the pods are ignored, and value false indicates that the pods are evicted. This configuration does not differentiate PVC types (local PVs or EVS). nodeFit: whether to consider the existing scheduling configurations such as node affinity and taint on the node during descheduling. Value true indicates that the existing scheduling configurations will be considered, and value false indicates that those will be ignored. priorityThreshold: priority setting. If the priority of a pod is greater than or equal to the value of this parameter, the pod will not be evicted. Example: { "value": 100 } Configurations for the HighNodeUtilization policy: evictableNamespaces: namespaces where the eviction policy takes effect. The default value is the namespaces other than kube-system. Example: { "exclude": ["kube-system"] } thresholds: threshold for evicting pods from a node. When the CPU or memory usage of a node is less than the threshold, the pods on the node will be evicted. Example: { "cpu": 25, "memory": 25 }

Click OK.

Use Cases

HighNodeUtilization

Check the nodes in a cluster. It is found that some nodes are under-utilized.
Edit the Volcano parameters to enable the descheduler and set the CPU and memory usage thresholds to 25. When the CPU and memory usage of a node is less than 25%, pods on the node will be evicted.
After the policy takes effect, pods on the node with IP address 192.168.44.152 will be migrated to the node with IP address 192.168.54.65 for minimized resource fragments.

Common Issues

If an input parameter is incorrect, for example, the entered value is beyond the accepted value range or in an incorrect format, an event will be generated.

Parent topic: Resource Usage-based Scheduling

Previous topic: Bin Packing

Next topic: Node Pool Affinity