Container Service for Kubernetes:Use PPUs in a Lingjun node pool

Prerequisites

1. An ACK managed Pro cluster is created, and the ack-rdma-device-plugin and ack-ppu-device-plugin add-ons are installed.

   When you install the ack-ppu-device-plugin add-on, select the Enable PPU Topology option. After this option is enabled, the add-on reports PPU topology information, and PPU topology-aware scheduling assigns the optimal PPU combination to your tasks based on the physical topology of PPUs by default. Scheduling requests for 1, 2, 4, 8, or 16 PPUs are supported.

2. Create a Lingjun node pool.

3. Add existing Lingjun nodes (PPU nodes) to the Lingjun node pool.

Verify that PPU nodes are ready

After you add PPU nodes to the Lingjun node pool, check whether the nodes are ready:

kubectl get nodes -l aliyun.accelerator/xpu_type=ppu

After a PPU node is ready, the node has the following PPU-related Kubernetes labels:

View the labels of a node:

kubectl get node <NODE_NAME> -o jsonpath='{.metadata.labels}' | jq 'to_entries[ ] | select(.key | startswith("aliyun.accelerator"))'

Enable the binpack scheduling policy

To prioritize placing PPU tasks on the same node and reduce resource fragmentation, it is recommended that you enable the binpack scheduling policy.

1. Log on to the ACK console. In the left navigation pane, click Clusters.

2. On the Clusters page, click the name of your cluster. In the left navigation pane, click Components and Add-ons .

3. Find Kube Scheduler and click Configuration. Configure the parameters in the Kube Scheduler Parameters dialog box.

   1. Enable Pod scheduling prioritizes consolidating Pods on fewer nodes.

   2. In the binpackResourceWeight area, click Add and set resourceName to alibabacloud.com/ppu to enable the binpack scheduling policy for PPU resources.

Deploy a PPU application

After the configuration is complete, request PPU resources by using resources.limits in a workload. The following example shows how to deploy an application that requests one PPU:

apiVersion: apps/v1
kind: Deployment
metadata:
  labels:
    app: ppu
  name: ppu-test
  namespace: default
spec:
  selector:
    matchLabels:
      app: ppu
  template:
    metadata:
      labels:
        app: ppu
    spec:
      containers:
        - image: anolis-registry.cn-zhangjiakou.cr.aliyuncs.com/openanolis/nginx:1.14.1-8.6
          imagePullPolicy: IfNotPresent
          name: ppu
          ports:
            - containerPort: 80
              protocol: TCP
          resources:
            limits:
              alibabacloud.com/ppu: "1"
      dnsPolicy: ClusterFirst
      restartPolicy: Always
      schedulerName: default-scheduler
      tolerations:
        - key: node-role.alibabacloud.com/lingjun
          operator: Exists
          effect: NoSchedule

Key configurations:

• resources.limits.alibabacloud.com/ppu: The number of requested PPUs. Valid values: 1, 2, 4, 8, and 16.

• tolerations: Add a toleration for the node-role.alibabacloud.com/lingjun taint on Lingjun nodes. Otherwise, pods cannot be scheduled to PPU nodes.

Verify the result

Check whether the pod is scheduled to a PPU node and running normally:

kubectl get pod -l app=ppu -o wide

In the expected output, the status of the pod is Running, and the NODE column shows the name of a PPU node.

View the PPU allocation of the node on which the pod runs:

kubectl describe node <NODE_NAME> | grep -A 5 "Allocated resources"

The expected output shows that the alibabacloud.com/ppu resource is allocated.