View node allocatable resources and node resource reservation policy-Container Service for Kubernetes(ACK)-阿里云帮助中心

ACK reserves node CPU and memory for Kubernetes and system components, which creates a gap between total node capacity (Capacity) and schedulable resources (Allocatable). ACK applies a default reservation policy and also supports custom reservations through kubelet configuration.

Limits

Custom resource reservation requires Kubernetes 1.20 or later. To upgrade, see Manually upgrade a cluster.

Scope of impact

Custom resource reservation and its scope of impact

To change reservation values, configure kubelet for a node pool. Changes apply immediately to existing nodes. New nodes—including those from scaling or Add Existing Node—also use the updated configuration.

Important

Do not manually edit the kubelet configuration file on the command line. Doing so may cause configuration conflicts and unexpected behavior during node pool O&M.
Increasing resource reservation reduces allocatable resources. On nodes with high resource usage, this may trigger pod eviction. Set values carefully.

Default resource reservation scope of impact

ACK may update default reservation values over time. Updated values apply automatically when you perform node-level changes—for example, upgrading the cluster or node pool, or modifying kubelet parameters. Without such changes, existing nodes retain previous values for stability.

View node allocatable resources

View a node’s total capacity and allocatable resources:

kubectl describe node [NODE_NAME] | grep Allocatable -B 7 -A 6

Expected output:

Capacity:
  cpu:                4                 # Total CPU cores on the node.
  ephemeral-storage:  123722704Ki       # Total ephemeral storage on the node, in KiB.
  hugepages-1Gi:      0
  hugepages-2Mi:      0
  memory:             7925980Ki         # Total memory on the node, in KiB.
  pods:               64
Allocatable:
  cpu:                3900m             # Allocatable CPU cores.
  ephemeral-storage:  114022843818      # Allocatable ephemeral storage, in bytes.
  hugepages-1Gi:      0
  hugepages-2Mi:      0
  memory:             5824732Ki         # Allocatable memory, in KiB.
  pods:               64

Calculate allocatable resources

Use this formula to calculate allocatable resources:Allocatable = Capacity − Reserved − Eviction threshold

Formula details:

Total resources correspond to the Capacity field in the node query output.
See Resource reservation policy details.
See Node-pressure eviction for eviction thresholds.

Resource reservation policy details

Reserved resources depend on several factors:

Higher-spec ECS instances run more pods, so ACK reserves more resources for Kubernetes components.
Windows nodes reserve more resources than Linux nodes for the OS and Windows Server components. See Create and manage Windows node pools.

ACK calculates reserved resources based on CPU and memory ranges—the total equals the sum across all ranges. In Kubernetes 1.28, ACK optimized the algorithm to reduce reservations. See Manually upgrade a cluster.

Reserved resources split equally between Kubernetes components (kubeReserved) and system processes (systemReserved). On a 4-core node, ACK reserves 80 millicores in Kubernetes 1.28+ (40 kubeReserved + 40 systemReserved), or 100 millicores in 1.20 through 1.27 (50 + 50).

CPU resource reservation policy

Kubernetes 1.28 and later

Total CPU reservation for compute nodes:

For a 32-core node, total CPU reservation is calculated as follows:

1000 × 6% + 1000 × 1% + 1000 × 2 × 0.5% + (32000 − 4000) × 0.25% = 150 millicores

Kubernetes 1.20 through 1.27

Total CPU reservation for compute nodes:

For a 32-core node, total CPU reservation is calculated as follows:

100 + (32000 − 4000) × 2.5% = 800 millicores

Memory resource reservation policy

Kubernetes 1.28 and later

Total memory reservation formula:

Total memory reservation = min(11 × ($max_num_pods) + 255, 25% × node memory). The final value is the smaller of 11 × ($max_num_pods) + 255 and 25% of node memory.

$max_num_pods: Maximum number of pods supported on the node.
Note
Maximum pod counts vary by network plug-in. See Maximum pods per node for details.

You can log on to the Container Service Management Console and view the maximum pod count on the Nodes page.
- Terway: Maximum pods per node = Maximum container-network pods + Host-network pods.
- Flannel: Set by you when creating the cluster.
Node memory: Actual usable memory, in MiB.

For example, with Terway (shared ENI, multiple IPs) on an ecs.g7.16xlarge (256 GiB memory): maximum pods = (8 − 1) × 30 + 3 = 213. Total memory reservation = min(11 × (210 + 3) + 255, 256 × 1024 × 25%) = 2598 MiB.

Kubernetes 1.20 through 1.27

Total memory reservation for a compute node:

For a node with 256 GiB of memory, total memory reservation is calculated as follows:

4 × 25% + (8 − 4) × 20% + (16 − 8) × 10% + (128 − 16) × 6% + (256 − 128) × 2% = 11.88 GiB

Example default resource reservations for ACK nodes

Note

See Instance families for ECS instance type details.

Total node resources				Reserved resources (Kubernetes 1.28 and later)		Reserved resources (Kubernetes 1.20 through 1.27)
Sample instance type	CPU (cores)	Memory (GiB)	Maximum Pods per Node (using the Terway shared ENI multi-IP mode as an example)	CPU (milicore)	Memory (MiB)	CPU (millicores)	Memory (MiB)
ECS c7 instance types	2	4	15	70	420	100	1024
	4	8	48	80	783	100	1843
	8	16	48	90	783	200	2662
	16	32	213	110	2598	400	3645
	32	64	213	150	2598	800	5611
	64	128	213	230	2598	1600	9543
	128	256	423	390	4908	2400	12164
ECS ebmc7a instance types	256	512	453	710	5238	3040	17407

FAQ

How do I view total CPU and memory on a node?

CPU

View total CPU cores:

cat /proc/cpuinfo | grep processor

Expected output:

processor       : 0
processor       : 1
processor       : 2
processor       : 3

Memory

View total memory:

cat /proc/meminfo | grep MemTotal

Expected output:

MemTotal:        7660952 kB

Why is available memory less than the instance type specification?

Instance type specs include total memory (including OS usage), so available memory is always less. See After purchasing an instance, why does the displayed memory differ from the instance type specification?.

References

Customize resource reservations and eviction thresholds by configuring kubelet for a node pool.
Total pod requests on a node must not exceed allocatable resources, or pods fail to schedule. Use ACK resource profiling to analyze historical usage and set accurate requests for native Kubernetes workloads. See Create a Deployment.