Cluster capacity analysis framework
Implementation of cluster capacity analysis.
Intro
As new pods get scheduled on nodes in a cluster, more resources get consumed.
Monitoring available resources in the cluster is very important
as operators can increase the current resources in time before all of them get exhausted.
Or, carry different steps that lead to increase of available resources.
Cluster capacity consists of capacities of individual cluster nodes.
Capacity covers CPU, memory, disk space and other resources.
Overall remaining allocatable capacity is a rough estimation since it does not assume all resources being distributed among nodes.
Goal is to analyze remaining allocatable resources and estimate available capacity that is still consumable
in terms of a number of instances of a pod with given requirements that can be scheduled in a cluster.
Build and Run
Build the framework:
$ make build
and run the analysis:
$ ./cluster-capacity --kubeconfig <path to kubeconfig> --podspec=examples/pod.yaml
For more information about available options run:
$ ./cluster-capacity --help
Demonstration
Assuming a cluster is running with 4 nodes and 1 master with each node with 2 CPUs and 4GB of memory.
With pod resource requirements to be 150m
of CPU and 100Mi
of Memory.
$ ./cluster-capacity --kubeconfig <path to kubeconfig> --podspec=pod.yaml --verbose
Pod requirements:
- cpu: 150m
- memory: 100Mi
The cluster can schedule 52 instance(s) of the pod.
Termination reason: FailedScheduling: pod (small-pod-52) failed to fit in any node
fit failure on node (kube-node-1): Insufficient cpu
fit failure on node (kube-node-4): Insufficient cpu
fit failure on node (kube-node-2): Insufficient cpu
fit failure on node (kube-node-3): Insufficient cpu
Pod distribution among nodes:
- kube-node-1: 13 instance(s)
- kube-node-4: 13 instance(s)
- kube-node-2: 13 instance(s)
- kube-node-3: 13 instance(s)
To decrease available resources in the cluster you can use provided RC (examples/rc.yml
):
$ kubectl create -f examples/rc.yml
E.g. to change a number of replicas to 6
, you can run:
$ kubectl patch -f examples/rc.yml -p '{"spec":{"replicas":6}}'
Once the number of running pods in the cluster grows and the analysis is run again,
the number of schedulable pods decreases as well:
$ ./cluster-capacity --kubeconfig <path to kubeconfig> --podspec=pod.yaml --verbose
Pod requirements:
- cpu: 150m
- memory: 100Mi
The cluster can schedule 46 instance(s) of the pod.
Termination reason: FailedScheduling: pod (small-pod-46) failed to fit in any node
fit failure on node (kube-node-1): Insufficient cpu
fit failure on node (kube-node-4): Insufficient cpu
fit failure on node (kube-node-2): Insufficient cpu
fit failure on node (kube-node-3): Insufficient cpu
Pod distribution among nodes:
- kube-node-1: 11 instance(s)
- kube-node-4: 12 instance(s)
- kube-node-2: 11 instance(s)
- kube-node-3: 12 instance(s)
cluster capacity
command has a flag --output (-o)
to format its output as json or yaml.
$ ./cluster-capacity --kubeconfig <path to kubeconfig> --podspec=pod.yaml -o json
$ ./cluster-capacity --kubeconfig <path to kubeconfig> --podspec=pod.yaml -o yaml
The json or yaml output is not versioned and is not guaranteed to be stable across various releases.
Running Cluster Capacity as a Job Inside of a Pod
Running the cluster capacity tool as a job inside of a pod has the advantage of
being able to be run multiple times without needing user intervention.
Follow these example steps to run Cluster Capacity as a job:
1. Create a Container that runs Cluster Capacity
In this example we create a simple Docker image utilizing the Dockerfile found in the root directory and tag it with cluster-capacity-image
:
$ docker build -t cluster-capacity-image .
2. Setup an authorized user with the necessary permissions
A. Create a role:
$ cat << EOF| kubectl create -f -
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1beta1
metadata:
name: cluster-capacity-role
rules:
- apiGroups: [""]
resources: ["pods", "nodes", "persistentvolumeclaims", "persistentvolumes", "services"]
verbs: ["get", "watch", "list"]
EOF
B. Create the service account which will be used to run the job:
$ kubectl create sa cluster-capacity-sa
C. Add the role to the service account:
$ kubectl create clusterrolebinding cluster-capacity-role \
--clusterrole=cluster-capacity-role \
--serviceaccount=default:cluster-capacity-sa
3. Define and create the pod specification (pod.yaml):
apiVersion: v1
kind: Pod
metadata:
name: small-pod
labels:
app: guestbook
tier: frontend
spec:
containers:
- name: php-redis
image: gcr.io/google-samples/gb-frontend:v4
imagePullPolicy: Always
resources:
limits:
cpu: 150m
memory: 100Mi
requests:
cpu: 150m
memory: 100Mi
The cluster capacity analysis is mounted in a volume using a
ConfigMap
named cluster-capacity-configmap
to mount input pod spec file
pod.yaml
into a volume test-volume
at the path /test-pod
.
$ kubectl create configmap cluster-capacity-configmap \
--from-file pod.yaml
4. Create the job specification (cluster-capacity-job.yaml):
apiVersion: batch/v1
kind: Job
metadata:
name: cluster-capacity-job
spec:
parallelism: 1
completions: 1
template:
metadata:
name: cluster-capacity-pod
spec:
containers:
- name: cluster-capacity
image: cluster-capacity-image
imagePullPolicy: "Never"
volumeMounts:
- mountPath: /test-pod
name: test-volume
env:
- name: CC_INCLUSTER
value: "true"
command:
- "/bin/sh"
- "-ec"
- |
/bin/cluster-capacity --podspec=/test-pod/pod.yaml --verbose
restartPolicy: "Never"
serviceAccountName: cluster-capacity-sa
volumes:
- name: test-volume
configMap:
name: cluster-capacity-configmap
Note the environment variable CC_INCLUSTER
the example above is required. This is used to indicate to the cluster capacity tool that it is running inside a cluster as a pod.
The pod.yaml
key of the ConfigMap
is the same as the pod specification file
name, though it is not required. By doing this, the input pod spec file can be
accessed inside the pod as /test-pod/pod.yaml
.
5. Run the cluster capacity image as a job in a pod:
$ kubectl create -f cluster-capacity-job.yaml
6. Check the job logs to find the number of pods that can be scheduled in the cluster:
$ kubectl logs jobs/cluster-capacity-job
small-pod pod requirements:
- CPU: 150m
- Memory: 100Mi
The cluster can schedule 52 instance(s) of the pod small-pod.
Termination reason: Unschedulable: No nodes are available that match all of the
following predicates:: Insufficient cpu (2).
Pod distribution among nodes:
small-pod
- 192.168.124.214: 26 instance(s)
- 192.168.124.120: 26 instance(s)
Pod spec generator: genpod
genpod
is an internal tool to cluster capacity, and could be used to create sample pod spec.
In general, users are recommended to provide their own pod spec file as part of analysis
As pods are part of a namespace with resource limits and additional constraints (e.g. node selector forced by namespace annotation),
it is natural to analyse how many instances of a pod with maximal resource requirements can be scheduled.
In order to generate the pod spec, you can run:
$ genpod --kubeconfig <path to kubeconfig> --namespace <namespace>
Assuming at least one resource limits object is available with at least one maximum resource type per pod.
If multiple resource limits objects per namespace are available, minimum of all maximum resources per type is taken.
If a namespace is annotated with openshift.io/node-selector
, the selector is set as pod's node selector.
Example:
Assuming cluster-capacity
namespace with openshift.io/node-selector: "region=hpc,load=high"
annotation
and resource limits are created (see examples/namespace.yml
and examples/limits.yml
)
$ kubectl describe limits hpclimits --namespace cluster-capacity
Name: hpclimits
Namespace: cluster-capacity
Type Resource Min Max Default Request Default Limit Max Limit/Request Ratio
---- -------- --- --- --------------- ------------- -----------------------
Pod cpu 10m 200m - - -
Pod memory 6Mi 100Mi - - -
Container memory 6Mi 20Mi 6Mi 6Mi -
Container cpu 10m 50m 10m 10m -
$ genpod --kubeconfig <path to kubeconfig> --namespace cluster-capacity
apiVersion: v1
kind: Pod
metadata:
creationTimestamp: null
name: cluster-capacity-stub-container
namespace: cluster-capacity
spec:
containers:
- image: gcr.io/google_containers/pause:2.0
imagePullPolicy: Always
name: cluster-capacity-stub-container
resources:
limits:
cpu: 200m
memory: 100Mi
requests:
cpu: 200m
memory: 100Mi
dnsPolicy: Default
nodeSelector:
load: high
region: hpc
restartPolicy: OnFailure
status: {}
Roadmap
Underway:
- analysis covering scheduler and admission controller
- generic framework for any scheduler created by the default scheduler factory
- continuous stream of estimations
Would like to get soon:
- include multiple schedulers
- accept a list (sequence) of pods
- extend analysis with volume handling
- define common interface each scheduler need to implement if embedded in the framework
Other possibilities:
- incorporate re-scheduler
- incorporate preemptive scheduling
- include more of Kubelet's behaviour (e.g. recognize memory pressure, secrets/configmap existence test)