Platform For AI:Use instance interconnection for distributed training

Prerequisites

• Create multiple instances from a general computing resource group or Lingjun resource group in the same VPC.

• Set the Internet access gateway for the resource group to Private Gateway.

• Place all instances in the same cluster. Lingjun instances and general computing resource instances cannot interconnect.

• Some instance types support Remote Direct Memory Access (RDMA) or enhanced RDMA (eRDMA). For more information, see Default variables (pre-configured by the platform) and Limitations.

  DSW and DLC provide the same RDMA/eRDMA features. For more information, see the DLC documentation.

Key features

• DSW provides pre-configured, high-performance network environment variables optimized for different resources and network architectures.

  • For Lingjun resource instances, see Default variables (pre-configured by the platform).

  • For general computing resource instances, see Platform-preconfigured environment variables.

• Use RDMA/eRDMA for interconnection on nodes that support it.

• Interconnect instances using their instance IDs as DNS domain names.

These features enable distributed task development and debugging across multiple machines and GPUs.

Steps

1. Use DSW instance cloning to start the required number of instances with identical environments.

2. (Optional) Install the RDMA/eRDMA library.

   1. For Lingjun resources, use an image that contains RDMA. For more information, see Configure an image.

   2. For general computing resources, follow the instructions in Install eRDMA library.

3. Test network connectivity by running ping with the target instance ID. For example: ping dsw-l28wnjdlyzj*********.

4. Configure and debug distributed tasks using your chosen distributed framework.