Extended clusters are an advanced disaster recovery architecture in Sangfor’s Hyper-Converged Infrastructure (HCI) platform, designed to achieve active-active data replication and high availability across data centers. It distributes the hosts of a hyper-converged cluster evenly across two separate data centers (referred to as the primary and backup fault domains). Through data replication technology and arbitration mechanisms, it ensures that if one data center fails, the business can automatically switch to the other data center.

The Stretch Cluster Network requirements are as follows:

1. Management Network Between Two Data Centers: The VXLAN and storage networks between the two data centers must be able to communicate with each other.
2. Storage Network Requirement: The network between storage networks must be 10GbE.
3. Storage Link Latency Between Data Centers: The storage link latency between data centers must be less than 5ms.
4. Witness Link Latency (Between Arbitration Node and Data Centers): The latency of the arbitration link must be less than 5ms.
5. Dual-active Cluster Extension Scenario (e.g., zero downtime requirement for business): The network latency in a dual-active extension cluster scenario must be less than 5ms.

The recommended network topology of the HCI stretched cluster is as follows.

Networking Description:

Machine Room 1 and machine Room 2

Machine room 1 and machine room 2 form an HCI cluster. Under the condition of ensuring the interconnection of the management network, VXLAN network, and storage network of the two machine rooms, configure corresponding redundancy strategies for the management network, VXLAN network, and storage network in the machine room. Stack the three networks with two switches and configure corresponding network aggregation on the HCI cluster. The recommendations are as follows:

Management network: Each node is aggregated with 2 * GE network interfaces and stacked with two separate management network switches.

VXLAN network: Each node is aggregated with 2 * GE network interfaces and stacked with two separate VXLAN network switches.

Storage network: Each node is aggregated with 2 * 10GE network interfaces and stacked with 2 separate storage network switches.

The two machine rooms can set their own Edge. It is also recommended that the edge network adopt 2 * GE network interfaces for aggregation to provide services to the external network. If the conditions allow, providing the individual edge to the two machine rooms is suggested. If the conditions do not allow it, set the edge in the main service's machine room.

Machine room network:

Machine room 1 and machine room 2 are directly connected through the second layer of the network. It is recommended to use a 10 Gigabit bare optical fiber network for aggregation, and the network latency in the machine room should be less than or equal to 1ms. The switches of the two machine rooms are stacked, and three different VLANs are divided on the stacked switches to carry the relevant data of the management network: the VXLAN network and the storage network of the two machine rooms.

Witness Node:

The witness node is important in the stretched cluster. It is different from the HCI system. To configure the witness node, you need to install the operating system of the witness node separately or purchase a separate aServer of the witness node. The regular communication between the witness node and the cluster shows whether the data copies in different computer rooms are abnormal and can be found in time. For the abnormal computer room, under the charge of the witness node, the business can continue to provide services in the computer room with normal data copies. The witness node must be deployed in a third-party site, the witness node needs to be able to communicate with the HCI cluster IP. The link delay between the witness node and the cluster shall not be less than 5ms, and the recommended bandwidth is 100m.