Node Resource Reservation Policy

Some node resources are used to run mandatory Kubernetes system components and resources to make the node as part of your cluster. Therefore, the total number of node resources and the number of allocatable node resources for your cluster are different. The larger the node specifications, the more the containers deployed on the node. Therefore, more node resources need to be reserved to run Kubernetes components.

To ensure node stability, a certain number of CCE node resources will be reserved for Kubernetes components (such as kubelet, kube-proxy, and docker) based on the node specifications.

CCE calculates the resources that can be allocated to user nodes as follows:

Allocatable resources = Total amount - Reserved amount - Eviction threshold

The memory eviction threshold is fixed at 100 MB.

Total amount indicates the available memory of the ECS, excluding the memory used by system components. Therefore, the total amount is slightly less than the memory of the node flavor. For details, see Why Is the Memory of an ECS Obtained by Running the free Command Inconsistent with the Actual Memory?

When the memory consumed by all pods on a node increases, the following behaviors may occur:

1. When the available memory of the node is lower than the eviction threshold, kubelet is triggered to evict the pod. For details about the eviction threshold in Kubernetes, see Node-pressure Eviction.
2. If a node triggers an OS memory insufficiency event (OOM) before kubelet reclaims memory, the system terminates the container. However, different from pod eviction, kubelet restarts the container based on the RestartPolicy of the pod.

Rules v1 for Reserving Node Memory

For clusters of versions earlier than v1.21.4-r0 and v1.23.3-r0, the v1 model is used for node memory reservation. For clusters of v1.21.4-r0, v1.23.3-r0, or later, the node memory reservation model is optimized to v2. For details, see Rules for Reserving Node Memory v2.

You can use the following formula calculate how much memory you should reserve for running containers on a node:

Total reserved amount = Reserved memory for system components + Reserved memory for kubelet to manage pods

Table 1 Reservation rules for system components

Total Memory (TM)	Reserved Memory for System Components
TM ≤ 8 GB	0 MB
8 GB < TM ≤ 16 GB	[(TM – 8 GB) x 1024 x 10%] MB
16 GB < TM ≤ 128 GB	[8 GB x 1024 x 10% + (TM – 16 GB) x 1024 x 6%] MB
TM > 128 GB	(8 GB x 1024 x 10% + 112 GB x 1024 x 6% + (TM – 128 GB) x 1024 x 2%) MB

Table 2 Reservation rules for kubelet

Total Memory (TM)	Number of Pods	Reserved Memory for kubelet
TM ≤ 2 GB	None	TM x 25%
TM > 2 GB	0 < Max. pods on a node ≤ 16	700 MB
	16 < Max. pods on a node ≤ 32	[700 + (Max. pods on a node – 16) x 18.75] MB
	32 < Max. pods on a node ≤ 64	[1024 + (Max. pods on a node – 32) x 6.25] MB
	64 < Max. pods on a node ≤ 128	[1230 + (Max. pods on a node – 64) x 7.80] MB
	Max. pods on a node > 128	[1740 + (Max. pods on a node – 128) x 11.20] MB

For a small-capacity node, adjust the maximum number of instances based on the site requirements. Alternatively, when creating a node on the CCE console, you can adjust the maximum number of instances for the node based on the node specifications.