Aim of this work will be to propose and develop a scheduling policy for intelligent placement of serverless workloads in different regions depending on the availability of renewable resources at given point of time. To achieve this, we will exploit kubernetes’ container orchestration capabilities and Knative’s capability of building serverless applications.
GreenCourier is a carbon-aware Kubernetes plugin to intelligently schedule serverless functions in regions of low carbon emission. GreenCourier optimises delivery of serverless functions across geo-spatial multi-cluster Kubernetes environment in the cloud for carbon efficiency. GreenCourier has production-ready tech stack and one-click away from integrating with existing geographically distributed clusters with Liqo.
We need a cluster with Knative enabled in management cluster and target clusters which are geographically distributed to be connected to management cluster using Liqo.
Knative can be installed using this guide or using the config YAMLs given in this repository under config/knative directory. It is important to note, a small change in service-core config specification as shown below:
apiVersion: v1
kind: ConfigMap
metadata:
name: config-features
namespace: knative-serving
labels:
app.kubernetes.io/name: knative-serving
app.kubernetes.io/component: controller
app.kubernetes.io/version: "1.7.1"
annotations:
knative.dev/example-checksum: "4d5feafc"
data:
kubernetes.podspec-schedulername: "enabled" # setting scheduler-name in knative spec should be enabled
Once Knative is deployed in Management cluster, it is important to setup multi-cluster topology using Liqo, depending on kubernetes distribution and providers, Liqo can be installed using following guide from official documentation.
Once Liqo is successfully installed on all Kubernetes clusters, we need to setup the topology within those clusters. To do that, every peer cluster should be run with following generate command:
liqoctl generate peer-command
Using the join-peer command is generated, which should be copied and should be executed in management cluster.
Once this is done for every peer cluster, we need to enable offloading in particular namespace, so that every pod is offloaded to corresponding peer clusters, using following command.
liqoctl offload namespace {$NAMESPACE} --namespace-mapping-strategy EnforceSameName --pod-offloading-strategy Remote
It is also important to annotate the peer clusters with corresponding region in management cluster, so that it is picked up by GreenCourier during scheduling. Annotating particular cluster is possible using following command after peering of cluster is done.
kubectl annotate node {$NODENAME} node.kubernetes.io/region={$REGION}
Once the cluster setup up is done, it is important for us to install metrics-collector in local cluster.
# Replace $USERNAME and $PASSWORD with WattTime credentials in deployment YAML
kubectl apply -f config/metrics-collector/metric-collector-deployment.yaml
Once metrics-collector is deployed, we can deploy the plugin code using Helm using following command:
helm install GreenCourier carbon-scheduler/charts/
When metrics-collector and plugin code is deployed, we can just start deploying functions in Knative with just addition of one line in the function spec YAML.
...
spec:
schedulerName: kube-carbon-scheduler
...
Example function spec with schedulerName added.
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: hello
namespace: default
spec:
template:
metadata:
annotations:
autoscaling.knative.dev/target: "10"
spec:
schedulerName: kube-carbon-scheduler
containers:
- image: gcr.io/knative-samples/helloworld-go
ports:
- containerPort: 8080
env:
- name: TARGET
value: "GreenCourier Demo"
Evaluation of GreenCourier was done against other scheduling schemes based on following three metrics.
- Pod placement efficiency
- Carbon Emission
- Scheduling and Response time
For functions whose execution time is less than 50ms:
For functions whose execution time is greater than 50ms:
Scheduling Latency:
Binding Latency:
