Help Center/ Cloud Container Engine/ User Guide (Kuala Lumpur Region)/ Best Practices/ Cluster/ Suggestions on CCE Cluster Selection
Updated on 2024-10-14 GMT+08:00
View PDF
Share

Suggestions on CCE Cluster Selection

When you use CCE to create a Kubernetes cluster, there are multiple configuration options and terms. This section compares the key configurations for CCE clusters and provides recommendations to help you create a cluster that better suits your needs.

Cluster Versions

Due to the fast iteration, many bugs are fixed and new features are added in the new Kubernetes versions. The old versions will be gradually eliminated. When creating a cluster, select the latest commercial version supported by CCE.

Network Models

This section describes the network models supported by CCE clusters. You can select one model based on your requirements.

After clusters are created, the network models cannot be changed. Exercise caution when selecting the network models.

Table 1 Network model comparison

Dimension

Tunnel Network

VPC Network

Application scenarios
  • Low requirements on performance: As the container tunnel network requires additional VXLAN tunnel encapsulation, it has about 5% to 15% of performance loss when compared with the other two container network models. Therefore, the container tunnel network applies to the scenarios that do not have high performance requirements, such as web applications, and middle-end and back-end services with a small number of access requests.
  • Large-scale networking: Different from the VPC network that is limited by the VPC route quota, the container tunnel network does not have any restriction on the infrastructure. In addition, the container tunnel network controls the broadcast domain to the node level. The container tunnel network supports a maximum of 2000 nodes.
  • High performance requirements: As no tunnel encapsulation is required, the VPC network model delivers the performance close to that of a VPC network when compared with the container tunnel network model. Therefore, the VPC network model applies to scenarios that have high requirements on performance, such as AI computing and big data computing.
  • Small- and medium-scale networks: Due to the limitation on VPC routing tables, it is recommended that the number of nodes in a cluster be less than or equal to 1000.

Core technology

OVS

IPvlan and VPC route

Applicable clusters

CCE standard cluster

CCE standard cluster

Container network isolation

Kubernetes native NetworkPolicy for pods

No

Interconnecting pods to a load balancer

Interconnected through a NodePort

Interconnected through a NodePort

Managing container IP addresses
  • Separate container CIDR blocks needed
  • Container CIDR blocks divided by node and dynamically added after being allocated
  • Separate container CIDR blocks needed
  • Container CIDR blocks divided by node and statically allocated (the allocated CIDR blocks cannot be changed after a node is created)

Network performance

Performance loss due to VXLAN encapsulation

No tunnel encapsulation, and cross-node traffic forwarded through VPC routers (The performance is so good that is comparable to that of the host network, but there is a loss caused by NAT.)

Networking scale

A maximum of 2000 nodes are supported.

Suitable for small- and medium-scale networks due to the limitation on VPC routing tables. It is recommended that the number of nodes be less than or equal to 1000.

Each time a node is added to the cluster, a route is added to the VPC routing tables. Evaluate the cluster scale that is limited by the VPC routing tables before creating the cluster.

Cluster CIDR Blocks

There are node CIDR blocks, container CIDR blocks, and Service CIDR blocks in CCE clusters. When planning network addresses, note that:

  • These three types of CIDR blocks cannot overlap with each other. Otherwise, a conflict will occur. All subnets (including those created from the secondary CIDR block) in the VPC where the cluster resides cannot conflict with the container and Service CIDR blocks.
  • There are sufficient IP addresses in each CIDR block.
    • The IP addresses in a node CIDR block must match the cluster scale. Otherwise, nodes cannot be created due to insufficient IP addresses.
    • The IP addresses in a container CIDR block must match the service scale. Otherwise, pods cannot be created due to insufficient IP addresses.

In complex scenarios, for example, multiple clusters use the same VPC or clusters are interconnected across VPCs, determine the number of VPCs, the number of subnets, the container CIDR blocks, and the communication modes of the Service CIDR blocks. For details, see Planning CIDR Blocks for a Cluster.

Service Forwarding Modes

kube-proxy is a key component of a Kubernetes cluster. It is responsible for load balancing and forwarding between a Service and its backend pod.

CCE supports the iptables and IPVS forwarding modes.

  • IPVS allows higher throughput and faster forwarding. It applies to scenarios where the cluster scale is large or the number of Services is large.
  • iptables is the traditional kube-proxy mode. This mode applies to the scenario where the number of Services is small or there are a large number of short concurrent connections on the client.

If high stability is required and the number of Services is less than 2000, the iptables forwarding mode is recommended. In other scenarios, the IPVS forwarding mode is recommended.

Node Specifications

The minimum specifications of a node are 2 vCPUs and 4 GiB memory. Evaluate based on service requirements before configuring the nodes. However, using many low-specification ECSs is not the optimal choice. The reasons are as follows:
  • The upper limit of network resources is low, which may result in a single-point bottleneck.
  • Resources may be wasted. If each container running on a low-specification node needs a lot of resources, the node cannot run multiple containers and there may be idle resources in it.
Advantages of using large-specification nodes are as follows:
  • The upper limit of the network bandwidth is high. This ensures higher resource utilization for high-bandwidth applications.
  • Multiple containers can run on the same node, and the network latency between containers is low.
  • The efficiency of pulling images is higher. This is because an image can be used by multiple containers on a node after being pulled once. Low-specifications ECSs cannot respond promptly because the images are pulled many times and it takes more time to scale these nodes.

Additionally, select a proper vCPU/memory ratio based on your requirements. For example, if a service container with large memory but fewer CPUs is used, configure the specifications with the vCPU/memory ratio of 1:4 for the node where the container resides to reduce resource waste.

Container Engines

CCE supports the containerd and Docker container engines. containerd is recommended for its shorter traces, fewer components, higher stability, and less consumption of node resources. Since Kubernetes 1.24, Dockershim is removed and Docker is no longer supported by default. For details, see Kubernetes is Moving on From Dockershim: Commitments and Next Steps. CCE clusters 1.27 do not support the Docker container engine.

Use containerd in typical scenarios. The Docker container engine is supported only in the following scenarios:

  • Docker in Docker (usually in CI scenarios)
  • Running the Docker commands on the nodes
  • Calling Docker APIs

Node OS

Service container runtimes share the kernel and underlying calls of nodes. To ensure compatibility, select a Linux distribution version that is the same as or close to that of the final service container image for the node OS.

Parent topic: Cluster