Coscheduling plugin

Based on the great work of Kubernetes new scheduler framework and kube-batch, it is very promising to impelment coscheduling feature(or named Gang scheduling). And this coscheduling plugin is for this purpose. It is just a POC(proof of concept) now. We will design and discuss more about it.

Design and implementation

It is implemented using PodGroup in kube-batch and permit plugin in Kubernetes new framework now. In the futher, we might re-design a new PodGroup that includes WaitTimeout and other attributes.

How to run

1. Run a kubernetes cluster

   We could run a Kubernetes cluster easily in a local machine. Or we could run this coschedling pluing in an existing Kuberntes cluster.

   # hack/local-up-cluster.sh        # run it in the Kuberentes repo directory
   

2. Install CRD PodGroup

   # kubectl apply -f https://raw.githubusercontent.com/kubernetes-sigs/kube-batch/master/deployment/kube-batch/templates/scheduling_v1alpha2_podgroup.yaml
   

3. Kill the default scheduler, and run the scheduler with coscheduling plugin

   # kill XXX   # XXX is the pid of the default scheduler
   # go build -o ./coscheduling ./main.go
   # ./coscheduling --v=5 --config ./hack/config.json
   

4. Test

   # kubectl create -f examples/podgroup.yaml
   # kubectl create -f ./examples/deployment.yaml
   

   In the scheduler log, we'll see logs as following:

   Wait for pod number of PodGroup to be 3, got 1 now
   ...
   Wait for pod number of PodGroup to be 3, got 2 now
   ...
   Wait for pod number of PodGroup to be 3, got 3 now
   

   When it reaches the min number 3, it will permit these waiting pods.

   We could also test other cases for it, will add more tests for it.