Batch-scheduler

Background

Currently, through the default scheduler of Kubernetes, we cannot ensure a group of pods scheduled at the same time . Under some scene, it would waste resources since some pods need work together, like spark, tensorflow and so on . So, batch-scheduler is aimed at solving the issue.

Method

Features

• lightweight
• no resource race
• gang scheduling

Implementation

Based on the latest scheduling framework, we designed the scheduler. So only one scheduler is needed to run in the cluster, which makes sure resource race would not happen.

This scheduler also makes sure gang scheduling, e.g.

• scene1

A group consists of 5 pods. The batch-scheduler would not schedule any pod until enough resources are found.

• scene2 Only 6 cpu exist in the cluster. Two groups require 5 cpus and 5 cpus are submitted, then only one and at least one group would be scheduled.

How to keep light-weight. We named a CRD PodGroup. When we would to running a group of pods, just need submit a PodGroup, e.g. group1 into the cluster. The pods needs to run as a group should only add a label named: group.batch .scheduler.tencent.com: group1

Main Progress

1. PreFilter: Compute resource requirements before we start predicts for a pod. If a pod is not permitted, we add it to freeze cache, then the pods belong to the same group would be rejected directly.

2. Less: this interface decides the sequence of pods. Currently, pods having higher Priority would be scheduled first. If pods have same the priority, PodGroup Creation time would be compared.

3. Permit: it is used for approving a pod or denying one. If a pod can be scheduled, but the number of pods belongs to the same group has not reached the min requested, it returns Wait. If a pod cannot be scheduled, it returns Unschedulable.

4. It is better to set MaxScheduleTime for a PodGroup. If one of the pods belong to the same PodGroup times out , other pods would also be rejected.

Build

# git clone git@github.com/tenstack/batch-scheduler.git
# make build

Deploy

• Deploy CRD

# cd deploy
# kubectl apply -f deploy crd.yaml

• Configuration

Default config has been written, but kube_config in it should be changed to your self stored.

• Deploy batch-scheduler

# cd deploy
# bash start.sh

Example

This example shows the resource race scene. Only 8 cpu exist in the cluster, and 0.9 has been occupied.

Allocated resources:
  (Total limits may be over 100 percent, i.e., overcommitted.)
  Resource           Requests    Limits
  --------           --------    ------
  cpu                900m (11%)  0 (0%)
  memory             140Mi (0%)  340Mi (2%)
  ephemeral-storage  0 (0%)      0 (0%)
Events:              <none>

• Yaml file named sts-group-valid-race.yaml is as follow

apiVersion: batch.scheduler.tencent.com/v1
kind: PodGroup
metadata:
  name: group1
  namespace: default
spec:
  minMember: 5
---
apiVersion: batch.scheduler.tencent.com/v1
kind: PodGroup
metadata:
  name: group2
  namespace: default
spec:
  minMember: 5
---
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: web-group-race1
spec:
  selector:
    matchLabels:
      app: nginx
  podManagementPolicy: Parallel
  serviceName: "nginx"
  replicas: 5
  template:
    metadata:
      labels:
        group.batch.scheduler.tencent.com: "group1"
        app: nginx
        type: node
    spec:
      containers:
      - name: nginx
        image: nginx
        ports:
        - containerPort: 80
          name: web
        resources:
          limits:
            cpu: "1"
          requests:
            cpu: "1"
---
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: web-group-race2
spec:
  selector:
    matchLabels:
      app: nginx
  podManagementPolicy: Parallel
  serviceName: "nginx"
  replicas: 5
  template:
    metadata:
      labels:
        group.batch.scheduler.tencent.com: "group2"
        app: nginx
        type: node
    spec:
      containers:
      - name: nginx
        image: nginx
        ports:
        - containerPort: 80
          name: web
        resources:
          limits:
            cpu: "1"
          requests:
            cpu: "1"

• Submit it

# kubectl apply -f sts-group-valid-race.yaml

• Results

[root@cwd-dev ~]# kubectl get pod
NAME                            READY   STATUS              RESTARTS   AGE
web-group-race1-0               0/1     ContainerCreating   0          16s
web-group-race1-1               0/1     ContainerCreating   0          16s
web-group-race1-2               0/1     ContainerCreating   0          16s
web-group-race1-3               1/1     Running             0          16s
web-group-race1-4               0/1     ContainerCreating   0          16s
web-group-race2-0               0/1     Pending             0          16s
web-group-race2-1               0/1     Pending             0          16s
web-group-race2-2               0/1     Pending             0          16s
web-group-race2-3               0/1     Pending             0          16s
web-group-race2-4               0/1     Pending             0          16s