Platform For AI:Plan Lingjun network resources

The following figure shows the relationships among basic modules such as Lingjun VPDs, Lingjun connections, and Lingjun hubs. For more information about the modules, see Manage Lingjun VPDs, Manage Lingjun connection instances, Manage Lingjun hubs, and Manage LNIs.

Plan the number of Lingjun VPDs

For more information about how to configure Lingjun VPDs and Lingjun subnets, see Manage Lingjun VPDs.

• Scenarios in which RDMA communication is required

  Remote direct memory access (RDMA) communication cannot be implemented across Lingjun VPDs. Therefore, Lingjun nodes that require RDMA communication must be deployed in the same Lingjun VPD.

• Scenarios in which RDMA communication is not required

  • If your business does not require isolation between Lingjun VPDs, we recommend that you use only one Lingjun VPD.

  • If your business scenario requires isolation between Lingjun VPDs, such as a scenario in which production and test environments are separated, we recommend that you use multiple Lingjun VPDs. In this case, you must take note of the following items:

    • You can deploy multiple Lingjun clusters in a Lingjun VPD, but a Lingjun cluster must reside in only one Lingjun VPD.

    • Lingjun nodes across Lingjun VPDs cannot form a Lingjun cluster because they cannot use RDMA communication.

    • We recommend that you plan the CIDR blocks of different Lingjun VPDs in advance in case of network exceptions caused by CIDR block conflicts.

    • After you associate multiple Lingjun VPDs with a Lingjun hub, you can adjust the routing policies in the Lingjun hub to control the connection or isolation between different Lingjun VPDs. For more information, see Manage Lingjun hubs.

Plan the number of Lingjun subnets

To prevent excessive Address Resolution Protocol (ARP) entries and the resource exhaustion on Lingjun nodes caused by a large number of Lingjun network interfaces (LNIs) in a Lingjun subnet, we recommend that you connect the LNIs of the same Lingjun node to different Lingjun subnets. This ensures stable RDMA communication and uninterrupted training tasks.

For example, if you want to configure five Lingjun LNIs for a Lingjun node, we recommend that you create five Lingjun subnets in a Lingjun VPD and separately connect the five LNIs to the five Lingjun subnets.

Plan CIDR blocks

When you create Lingjun VPDs and Lingjun subnets, you must specify a private IP address range for them in the form of a CIDR block.

• Plan Lingjun VPDs

  You cannot use 100.64.0.0/10, 224.0.0.0/4, 127.0.0.0/8, or 169.254.0.0/16 as the CIDR block of a Lingjun VPD.

  Note

  • You must plan a Lingjun VPD in advance. The Lingjun VPD cannot conflict with the CIDR blocks of other networks to which the Lingjun cluster is to be connected, such as CIDR blocks of VPCs or data centers.

  • The number of available IP addresses in a Lingjun VPD determines the maximum number of nodes that can be deployed in the Lingjun cluster. You must prepare a CIDR block whose subnet mask length is greater than 22 bits to prevent the cluster from being unable to be scaled up.

• Plan Lingjun subnets

  The CIDR block of a Lingjun subnet must be a subset of the Lingjun VPD to which the Lingjun subnet belongs. For example, if the CIDR block of a Lingjun VPD is 192.168.0.0/16, the CIDR block of a Lingjun subnet that belongs to this VPD ranges from 192.168.0.0/17 to 192.168.0.0/29.

  Note

  • The Lingjun subnet mask must be no less than 23 bits in length, which means that the number of IP addresses in a Lingjun subnet cannot exceed 512. This prevents the resource exhaustion on Lingjun nodes caused by excessive ARP entries.

  • When you plan the CIDR blocks of a Lingjun subnet, you must take into account the number of IP addresses of Lingjun nodes in the Lingjun subnet.