Kubernetes nodes typically run many OS system daemons in addition to kubernetes daemons like kubelet, runtime, etc. and user pods. Kubernetes assumes that all the compute resources available, referred to as Capacity, in a node are available for user pods. In reality, system daemons use non-trivial amount of resources and their availability is critical for the stability of the system. To address this issue, this proposal introduces the concept of Allocatable which identifies the amount of compute resources available to user pods. Specifically, the kubelet will provide a few knobs to reserve resources for OS system daemons and kubernetes daemons.

By explicitly reserving compute resources, the intention is to avoid overcommiting the node and not have system daemons compete with user pods. The resources available to system daemons and user pods will be capped based on user specified reservations.

If Allocatable is available, the scheduler will use that instead of Capacity, thereby not overcommiting the node.

1. Node Capacity - Already provided as NodeStatus.Capacity, this is total capacity read from the node instance, and assumed to be constant.
2. System-Reserved (proposed) - Compute resources reserved for processes which are not managed by Kubernetes. Currently this covers all the processes lumped together in the /system raw container.
3. Kubelet Allocatable - Compute resources available for scheduling (including scheduled & unscheduled resources). This value is the focus of this proposal. See below for more details.
4. Kube-Reserved (proposed) - Compute resources reserved for Kubernetes components such as the docker daemon, kubelet, kube proxy, etc.

Add Allocatable (4) to NodeStatus:

type NodeStatus struct {
  ...
  // Allocatable represents schedulable resources of a node.
  Allocatable ResourceList `json:"allocatable,omitempty"`
  ...
}

Allocatable will be computed by the Kubelet and reported to the API server. It is defined to be:

   [Allocatable] = [Node Capacity] - [Kube-Reserved] - [System-Reserved] - [Hard-Eviction-Threshold]

The scheduler will use Allocatable in place of Capacity when scheduling pods, and the Kubelet will use it when performing admission checks.

Note: Since kernel usage can fluctuate and is out of kubernetes control, it will be reported as a separate value (probably via the metrics API). Reporting kernel usage is out-of-scope for this proposal.

KubeReserved is the parameter specifying resources reserved for kubernetes components (4). It is provided as a command-line flag to the Kubelet at startup, and therefore cannot be changed during normal Kubelet operation (this may change in the future).

The flag will be specified as a serialized ResourceList, with resources defined by the API ResourceName and values specified in resource.Quantity format, e.g.:

--kube-reserved=cpu=500m,memory=5Mi

Initially we will only support CPU and memory, but will eventually support more resources like local storage and io proportional weights to improve node reliability.

In the initial implementation, SystemReserved will be functionally equivalent to KubeReserved, but with a different semantic meaning. While KubeReserved designates resources set aside for kubernetes components, SystemReserved designates resources set aside for non-kubernetes components (currently this is reported as all the processes lumped together in the /system raw container on non-systemd nodes).

To improve the reliability of nodes, kubelet evicts pods whenever the node runs out of memory or local storage. Together, evictions and node allocatable help improve node stability.

As of v1.5, evictions are based on overall node usage relative to Capacity. Kubelet evicts pods based on QoS and user configured eviction thresholds. More details in this doc

From v1.6, if Allocatable is enforced by default across all pods on a node using cgroups, pods cannot exceed Allocatable. Memory and CPU limits are enforced using cgroups, but there exists no easy means to enforce storage limits though. Enforcing storage limits using Linux Quota is not possible since it's not hierarchical. Once storage is supported as a resource for Allocatable, Kubelet has to perform evictions based on Allocatable in addition to Capacity.

Note that eviction limits are enforced on pods only and system daemons are free to use any amount of resources unless their reservations are enforced.

Here is an example to illustrate Node Allocatable for memory:

Node Capacity is 32Gi, kube-reserved is 2Gi, system-reserved is 1Gi, eviction-hard is set to <100Mi

For this node, the effective Node Allocatable is 28.9Gi only; i.e. if kube and system components use up all their reservation, the memory available for pods is only 28.9Gi and kubelet will evict pods once overall usage of pods crosses that threshold.

If we enforce Node Allocatable (28.9Gi) via top level cgroups, then pods can never exceed 28.9Gi in which case evictions will not be performed unless kernel memory consumption is above 100Mi.

In order to support evictions and avoid memcg OOM kills for pods, we will set the top level cgroup limits for pods to be Node Allocatable + Eviction Hard Thresholds.

However, the scheduler is not expected to use more than 28.9Gi and so Node Allocatable on Node Status will be 28.9Gi.

If kube and system components do not use up all their reservation, with the above example, pods will face memcg OOM kills from the node allocatable cgroup before kubelet evictions kick in. To better enforce QoS under this situation, Kubelet will apply the hard eviction thresholds on the node allocatable cgroup as well, if node allocatable is enforced. The resulting behavior will be the same for user pods. With the above example, Kubelet will evict pods whenever pods consume more than 28.9Gi which will be <100Mi from 29Gi which will be the memory limits on the Node Allocatable cgroup.

System daemons are expected to be treated similar to Guaranteed pods. System daemons can burst within their bounding cgroups and this behavior needs to be managed as part of kubernetes deployment. For example, Kubelet can have its own cgroup and share KubeReserved resources with the Container Runtime. However, Kubelet cannot burst and use up all available Node resources if KubeReserved is enforced.

Users are advised to be extra careful while enforcing SystemReserved reservation since it can lead to critical services being CPU starved or OOM killed on the nodes. The recommendation is to enforce SystemReserved only if a user has profiled their nodes exhaustively to come up with precise estimates.

To begin with enforce Allocatable on pods only. Once adequate monitoring and alerting is in place to track kube daemons, attempt to enforce KubeReserved based on heuristics. More on this in Phase 2.

The resource requirements of kube system daemons will grow over time as more and more features are added. Over time, the project will attempt to bring down utilization, but that is not a priority as of now. So expect a drop in Allocatable capacity over time.

Systemd-logind places ssh sessions under /user.slice. Its usage will not be accounted for in the nodes. Take into account resource reservation for /user.slice while configuring SystemReserved. Ideally /user.slice should reside under SystemReserved top level cgroup.

Following is the recommended cgroup configuration for Kubernetes nodes. All OS system daemons are expected to be placed under a top level SystemReserved cgroup. Kubelet and Container Runtime are expected to be placed under KubeReserved cgroup. The reason for recommending placing the Container Runtime under KubeReserved is as follows:

1. A container runtime on Kubernetes nodes is not expected to be used outside of the Kubelet.
2. It's resource consumption is tied to the number of pods running on a node.

Note that the hierarchy below recommends having dedicated cgroups for kubelet and the runtime to individually track their usage.

/ (Cgroup Root)
.
+..systemreserved or system.slice (Specified via `--system-reserved-cgroup`; `SystemReserved` enforced here *optionally* by kubelet)
.        .    .tasks(sshd,udev,etc)
.
.
+..podruntime or podruntime.slice (Specified via `--kube-reserved-cgroup`; `KubeReserved` enforced here *optionally* by kubelet)
.	 .
.	 +..kubelet
.	 .   .tasks(kubelet)
.        .
.	 +..runtime
.	     .tasks(docker-engine, containerd)
.	 
.
+..kubepods or kubepods.slice (Node Allocatable enforced here by Kubelet)
.	 .
.	 +..PodGuaranteed
.	 .	  .
.	 .	  +..Container1
.	 .	  .        .tasks(container processes)
.	 .	  .
.	 .        +..PodOverhead
.	 .        .        .tasks(per-pod processes)
.	 .        ...
.	 .
.	 +..Burstable
.	 .	  .
.	 .	  +..PodBurstable
.	 .	  .	    .
.	 .	  .	    +..Container1
.	 .	  .	    .         .tasks(container processes)
.	 .	  .	    +..Container2
.	 .	  .	    .         .tasks(container processes)
.	 .	  .	    .
.      	 .     	  .    	    ...
.	 .	  .
.	 .	  ...
.	 .
. 	 .
.	 +..Besteffort
.	 .	  .
.	 .	  +..PodBesteffort
.	 .	  .    	    .
.	 .	  .	    +..Container1
.	 .	  .	    .         .tasks(container processes)
.	 .	  .	    +..Container2
.	 .	  .	    .         .tasks(container processes)
.	 .	  .	    .
.      	 .     	  .    	    ...
.	 .	  .
. 	 .	  ...

systemreserved & kubereserved cgroups are expected to be created by users. If Kubelet is creating cgroups for itself and docker daemon, it will create the kubereserved cgroups automatically.

kubepods cgroups will be created by kubelet automatically if it is not already there. Creation of kubepods cgroup is tied to QoS Cgroup support which is controlled by --cgroups-per-qos flag. If the cgroup driver is set to systemd then Kubelet will create a kubepods.slice via systemd. By default, Kubelet will mkdir /kubepods cgroup directly via cgroupfs.

If Kubelet is managed using a container runtime, have the runtime create cgroups for kubelet under kubereserved.

Kubelet identifies it's own cgroup and exposes it's usage metrics via the Summary metrics API (/stats/summary) With docker runtime, kubelet identifies docker runtime's cgroups too and exposes metrics for it via the Summary metrics API. To provide a complete overview of a node, Kubelet will expose metrics from cgroups enforcing SystemReserved, KubeReserved & Allocatable too.

Status: Implemented v1.2

In this phase, Kubelet will support specifying KubeReserved & SystemReserved resource reservations via kubelet flags. The defaults for these flags will be "", meaning zero cpu or memory reservations. Kubelet will compute Allocatable and update Node.Status to include it. The scheduler will use Allocatable instead of Capacity if it is available.

Status: Targeted for v1.6

In this phase, Kubelet will automatically create a top level cgroup to enforce Node Allocatable across all user pods. The creation of this cgroup is controlled by --cgroups-per-qos flag.

Kubelet will support specifying the top level cgroups for KubeReserved and SystemReserved and support optionally placing resource restrictions on these top level cgroups.

Users are expected to specify KubeReserved and SystemReserved based on their deployment requirements.

Resource requirements for Kubelet and the runtime is typically proportional to the number of pods running on a node. Once a user identified the maximum pod density for each of their nodes, they will be able to compute KubeReserved using this performance dashboard. This blog post explains how the dashboard has to be interpreted. Note that this dashboard provides usage metrics for docker runtime only as of now.

Support for evictions based on Allocatable will be introduced in this phase.

New flags introduced in this phase are as follows:

1. --enforce-node-allocatable=[pods][,][kube-reserved][,][system-reserved]

   • This flag will default to pods in v1.6.
   • This flag will be a no-op unless --kube-reserved and/or --system-reserved has been specified.
   • If --cgroups-per-qos=false, then this flag has to be set to "". Otherwise its an error and kubelet will fail.
   • It is recommended to drain and restart nodes prior to upgrading to v1.6. This is necessary for --cgroups-per-qos feature anyways which is expected to be turned on by default in v1.6.
   • Users intending to turn off this feature can set this flag to "".
   • Specifying kube-reserved value in this flag is invalid if --kube-reserved-cgroup flag is not specified.
   • Specifying system-reserved value in this flag is invalid if --system-reserved-cgroup flag is not specified.
   • By including kube-reserved or system-reserved in this flag's value, and by specifying the following two flags, Kubelet will attempt to enforce the reservations specified via --kube-reserved & system-reserved respectively.

2. --kube-reserved-cgroup=<absolute path to a cgroup>

   • This flag helps kubelet identify the control group managing all kube components like Kubelet & container runtime that fall under the KubeReserved reservation.
   • Example: /kube.slice. Note that absolute paths are required and systemd naming scheme isn't supported.

3. --system-reserved-cgroup=<absolute path to a cgroup>

   • This flag helps kubelet identify the control group managing all OS specific system daemons that fall under the SystemReserved reservation.
   • Example: /system.slice. Note that absolute paths are required and systemd naming scheme isn't supported.

4. --experimental-node-allocatable-ignore-eviction-threshold

   • This flag is provided as an opt-out option to avoid including Hard eviction thresholds in Node Allocatable which can impact existing clusters.
   • The default value is false.

This phase is expected to improve Kubernetes node stability. However it requires users to specify non-default values for --kube-reserved & --system-reserved flags though.

The rollout of this phase has been long due and hence we are attempting to include it in v1.6.

Since KubeReserved and SystemReserved continue to have "" as defaults, the node's Allocatable does not change automatically. Since this phase requires node drains (or pod restarts/terminations), it is considered disruptive to users.

To rollback this phase, set --enforce-node-allocatable flag to "" and --experimental-node-allocatable-ignore-eviction-threshold to true. The former disables Node Allocatable enforcement on all pods and the latter avoids including hard eviction thresholds in Node Allocatable.

This rollout in v1.6 might cause the following symptoms:

1. If --kube-reserved and/or --system-reserved flags are also specified, OOM kills of containers and/or evictions of pods. This can happen primarily to Burstable and BestEffort pods since they can no longer use up all the resource available on the node.
2. Total allocatable capacity in the cluster reduces resulting in pods staying Pending because Hard Eviction Thresholds are included in Node Allocatable.

02/14/2017 - Discuss the rollout plan in sig-node meeting
02/15/2017 - Flip the switch to enable pod level cgroups by default
02/21/2017 - Merge phase 2 implementation
02/27/2017 - Kubernetes Feature complete (i.e. code freeze)
03/01/2017 - Send an announcement to kubernetes-dev@ about this rollout along with rollback options and potential issues. Recommend users to set kube and system reserved.
03/22/2017 - Kubernetes 1.6 release

Status: Targeted for v1.7

In this phase, Kubelet will expose usage metrics for KubeReserved, SystemReserved and Allocatable top level cgroups via Summary metrics API. Storage will also be introduced as a reservable resource in this phase.

Solution: Initially, do nothing (best effort). Let the kubernetes daemons overflow the reserved resources and hope for the best. If the node usage is less than Allocatable, there will be some room for overflow and the node should continue to function. If the node has been scheduled to allocatable (worst-case scenario) it may enter an unstable state, which is the current behavior in this situation.

A recommended alternative is to enforce KubeReserved once Kubelet supports it (Phase 2). In the future we may set a parent cgroup for kubernetes components, with limits set according to KubeReserved.

The community should be notified that an update to schedulers is recommended, but if a scheduler is not updated it falls under the above case of "scheduler is not allocatable-resources aware".