NUMA Affinity Scheduling

Background

When a node runs many CPU-bound pods, the workload can move to different CPU cores depending on whether the pod is throttled and which CPU cores are available at scheduling time. Many workloads are not sensitive to this migration and work fine without any intervention. However, in workloads where CPU cache affinity and scheduling latency significantly affect workload performance, additional latency will occur when CPU cores are from different NUMA nodes. To resolve this issue, kubelet allows you to use Topology Manager to replace the CPU management policies to determine node allocation.

Both the CPU Manager and Topology Manager are kubelet components, but they have the following limitations:

• The scheduler is not topology-aware. Therefore, the workload may be scheduled on a node and then fail on the node due to the Topology Manager. This is unacceptable for TensorFlow jobs. If any worker or ps failed on node, the job will fail.
• The managers are node-level that results in an inability to match the best node for NUMA topology in the whole cluster.

Volcano targets to lift the limitation to make scheduler NUMA topology aware so that:

• Pods are not scheduled to the nodes that NUMA topology does not match.
• Pods are scheduled to the most suitable node for NUMA topology.

For more information, see https://github.com/volcano-sh/volcano/blob/master/docs/design/numa-aware.md.

Application Scope

• CPU resource topology scheduling
• Pod-level topology policies

Pod Scheduling Prediction

After a topology policy is configured for pods, Volcano predicts matched nodes based on the topology policy. The scheduling process is as follows:

1. Volcano filters nodes with the same policy based on the topology policy configured for pods. The topology policy provided by Volcano is the same as that provided by the topology manager.
2. Among the nodes where the same policy applies, Volcano selects the nodes whose CPU topology meets the policy requirements for scheduling.

For example, two NUMA nodes provide resources, each with a total of 32 CPU cores. The following table lists resource allocation.

Figure 1 shows the scheduling of a pod after a topology policy is configured.

• When 9 CPU cores are requested by a pod and the best-effort topology policy is used, Volcano selects node 1 whose topology policy is also best-effort, and this policy allows the pod to be scheduled to multiple NUMA nodes. Therefore, the requested 9 CPU cores will be allocated to two NUMA nodes, and the pod can be scheduled to node 1.
• When 9 CPU cores are requested by a pod and the restricted topology policy is used, Volcano selects nodes 2 and 3 whose topology policy is also restricted, and each node provides a total of 9 CPU cores. However, the remaining CPU cores on node 2 or 3 are less than the requested. Therefore, the pod cannot be scheduled.
• When 17 CPU cores are requested by a pod and the restricted topology policy is used, Volcano selects nodes 2 and 3 whose topology policy is also restricted, this policy allows the pod to be scheduled to multiple NUMA nodes, and the upper CPU limit of both the nodes is less than 17. Then, the pod can be scheduled to node 3.
• When 17 CPU cores are requested by a pod and the single-numa-node topology policy is used, Volcano selects nodes whose topology policy is also single-numa-node. However, no node can provide a total of 17 CPU cores. Therefore, the pod cannot be scheduled.

Figure 1 Comparison of NUMA scheduling policies

Scheduling Priority

A topology policy aims to schedule pods to the optimal node. In this example, each node is scored to sort out the optimal node.

Principle: Schedule pods to the worker nodes that require the fewest NUMA nodes.

The scoring formula is as follows:

score = weight x (100 - 100 x numaNodeNum/maxNumaNodeNum)

Parameters:

• weight: the weight of NUMA Aware Plugin.
• numaNodeNum: the number of NUMA nodes required for running the pod on worker nodes.
• maxNumaNodeNum: the maximum number of NUMA nodes required for running the pod among all worker nodes.

For example, three nodes meet the CPU topology policy for a pod and the weight of NUMA Aware Plugin is set to 10.

• Node A: One NUMA node provides the CPU resources required by the pod (numaNodeNum = 1).
• Node B: Two NUMA nodes provide the CPU resources required by the pod (numaNodeNum = 2).
• Node C: Four NUMA nodes provide the CPU resources required by the pod (numaNodeNum = 4).

According to the preceding formula, maxNumaNodeNum is 4.

• score(Node A) = 10 x (100 - 100 x 1/4) = 750
• score(Node B) = 10 x (100 - 100 x 2/4) = 500
• score(Node C) = 10 x (100 - 100 x 4/4) = 0

Therefore, the optimal node is Node A.

Enabling NUMA Affinity Scheduling for Volcano

1. Enable static CPU management. For details, see Enabling the CPU Management Policy.
2. Configure a CPU topology policy.
   1. Log in to the CCE console and click the cluster name to access the cluster console. In the navigation pane, choose Nodes. On the right of the page, click the Node Pools tab and choose More > Manage in the Operation column of the target node pool.
   2. Change the kubelet Topology Management Policy (topology-manager-policy) value to the required CPU topology policy.

      Valid topology policies include none, best-effort, restricted, and single-numa-node. For details, see Pod Scheduling Prediction.

3. Enable the numa-aware add-on and the resource_exporter function.

   Volcano 1.7.1 or later

   1. Log in to the CCE console and click the cluster name to access the cluster console. In the navigation pane, choose Add-ons. On the right of the page, locate the Volcano add-on and click Edit. In the Parameters area, configure Volcano scheduler parameters.

      ...
          "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"
                          },
                          {
                              // add this also enable resource_exporter
                              "name": "numa-aware", 
                              // the weight of the NUMA Aware Plugin
                              "arguments": { 
                                 "weight": "10"
                              }
                          }
                      ]
                  },
                  {
                      "plugins": [
                          {
                              "name": "nodelocalvolume"
                          },
                          {
                              "name": "nodeemptydirvolume"
                          },
                          {
                              "name": "nodeCSIscheduling"
                          },
                          {
                              "name": "networkresource"
                          }
                      ]
                  }
              ]
          },
      ...

   Volcano earlier than 1.7.1

   1. The resource_exporter_enable parameter is enabled for the Volcano add-on to collect node NUMA information.

      {
         "plugins": {
            "eas_service": {
               "availability_zone_id": "",
               "driver_id": "",
               "enable": "false",
               "endpoint": "",
               "flavor_id": "",
               "network_type": "",
               "network_virtual_subnet_id": "",
               "pool_id": "",
               "project_id": "",
               "secret_name": "eas-service-secret"
            }
         },
         "resource_exporter_enable": "true"
      }

      After this function is enabled, you can view the NUMA topology information of the current node.

      kubectl get numatopo 
      NAME              AGE
      node-1            4h8m
      node-2            4h8m
      node-3            4h8m

   2. Enable the Volcano numa-aware algorithm add-on.
      kubectl edit cm -n kube-system volcano-scheduler-configmap

      kind: ConfigMap
      apiVersion: v1
      metadata:
        name: volcano-scheduler-configmap
        namespace: kube-system
      data:
        default-scheduler.conf: |-
          actions: "allocate, backfill, preempt"
          tiers:
          - plugins:
            - name: priority
            - name: gang
            - name: conformance
          - plugins:
            - name: overcommit
            - 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
              arguments:
                NetworkType: vpc-router
            - name: numa-aware # add it to enable numa-aware plugin
              arguments:
                weight: 10 # the weight of the NUMA Aware Plugin

Using Volcano to Configure NUMA Affinity Scheduling

1. Refer to the following examples for configuration.
   1. Example 1: Configure NUMA affinity for a Deployment.

      kind: Deployment
      apiVersion: apps/v1
      metadata:
        name: numa-tset
      spec:
        replicas: 1
        selector:
          matchLabels:
            app: numa-tset
        template:
          metadata:
            labels:
              app: numa-tset
            annotations:
              volcano.sh/numa-topology-policy: single-numa-node    # Configure the topology policy.
          spec:
            containers:
              - name: container-1
                image: nginx:alpine
                resources:
                  requests:
                    cpu: 2           # The value must be an integer and must be the same as that in limits.
                    memory: 2048Mi
                  limits:
                    cpu: 2           # The value must be an integer and must be the same as that in requests.
                    memory: 2048Mi
            imagePullSecrets:
            - name: default-secret

   2. Example 2: Create a Volcano job and enable NUMA affinity for it.

      apiVersion: batch.volcano.sh/v1alpha1
      kind: Job
      metadata:
        name: vj-test
      spec:
        schedulerName: volcano
        minAvailable: 1
        tasks:
          - replicas: 1
            name: "test"
            topologyPolicy: best-effort   # set the topology policy for task 
            template:
              spec:
                containers:
                  - image: alpine
                    command: ["/bin/sh", "-c", "sleep 1000"]
                    imagePullPolicy: IfNotPresent
                    name: running
                    resources:
                      limits:
                        cpu: 20
                        memory: "100Mi"
                restartPolicy: OnFailure

2. Analyze NUMA scheduling.

   The following table shows example NUMA nodes.

   Worker Node	Topology Manager Policy	Allocatable CPU Cores on NUMA Node 0	Allocatable CPU Cores on NUMA Node 1
   Node 1	single-numa-node	16	16
   Node 2	best-effort	16	16
   Node 3	best-effort	20	20

   In the preceding examples,

   • In example 1, 2 CPU cores are requested by a pod, and the single-numa-node topology policy is used. Therefore, the pod will be scheduled to node 1 with the same policy.
   • In example 2, 20 CPU cores are requested by a pod, and the best-effort topology policy is used. The pod will be scheduled to node 3 because it can allocate all the requested 20 CPU cores onto one NUMA node, while node 2 can do so on two NUMA nodes.

Checking NUMA Node Usage

Run the lscpu command to check the CPU usage of the current node.

# Check the CPU usage of the current node.
lscpu
...
CPU(s):              32
NUMA node(s):        2
NUMA node0 CPU(s):   0-15
NUMA node1 CPU(s):   16-31

Then, check the NUMA node usage.

# Check the CPU allocation of the current node.
cat /var/lib/kubelet/cpu_manager_state
{"policyName":"static","defaultCpuSet":"0,10-15,25-31","entries":{"777870b5-c64f-42f5-9296-688b9dc212ba":{"container-1":"16-24"},"fb15e10a-b6a5-4aaa-8fcd-76c1aa64e6fd":{"container-1":"1-9"}},"checksum":318470969}

The preceding example shows that two containers are running on the node. One container uses CPU cores 1 to 9 of NUMA node 0, and the other container uses CPU cores 16 to 24 of NUMA node 1.