多实例调度器

背景

目前单调度器在某些场景下无法满足高吞吐量的需求。除了针对单个调度器的性能优化之外，另一种选择是部署多个火山调度器来提高整体调度吞吐量。

解决

之前我们使用label将集群节点划分为多个section，每个火山调度器负责一个section，然后在Pod Spec中指定schedulerName并提交。这在某些情况下是不方便的，尤其是对于大型集群。该文档为用户提供了另一个选项来部署多个调度程序，它需要对工作负载和节点进行较少的修改。statefulset 用于部署火山调度程序。作业和节点根据哈希算法自动分配给调度程序。

通过statefulset启动多个实例， 由于statefulset里面每个POD都有独立id， ID 0 ~ replica -1, 便于实例获取id，进行任务切分。

kind: StatefulSet
apiVersion: apps/v1
metadata:
  name: volcano-scheduler
  namespace: volcano-system
  labels:
    app: volcano-scheduler
spec:
  replicas: 3
  selector:
    matchLabels:
      app: volcano-scheduler
  serviceName: "volcano-scheduler"
  template:
    metadata:
      labels:
        app: volcano-scheduler
    spec:
      serviceAccount: volcano-scheduler
      containers:
        - name: volcano-scheduler
          image: volcanosh/vc-scheduler:ae78900d21dce8522eb04b6817aac66c9abd01e2
          args:
            - --logtostderr
            - --scheduler-conf=/volcano.scheduler/volcano-scheduler.conf
            - -v=3
            - 2>&1
          imagePullPolicy: "IfNotPresent"
          env:
            - name: MULTI_SCHEDULER_ENABLE
              value: "true"
            - name: SCHEDULER_NUM
              value: "3"
            - name: SCHEDULER_POD_NAME
              valueFrom:
                fieldRef:
                  fieldPath: metadata.name
          volumeMounts:
            - name: scheduler-config
              mountPath: /volcano.scheduler
      volumes:
        - name: scheduler-config
          configMap:
            name: volcano-scheduler-configmap
---
apiVersion: v1
kind: Service
metadata:
  name: volcano-scheduler
  labels:
    app: volcano-scheduler
spec:
  ports:
  - port: 80
    name: volcano-scheduler
  clusterIP: None
  selector:
    app: volcano-scheduler

注意： SCHEDULER_NUM 需要和replica一致