Environment Dependencies

Software Dependencies

Ascend Docker Runtime

• The Docker version in the current environment must be 18.09 and later.

• The host machine must have the drivers and firmware installed. For details, see "Installing the NPU Driver and Firmware" in the CANN Software Installation Guide (Commercial) or "Installing the NPU Driver and Firmware" in the CANN Software Installation Guide (Community).

• For Atlas 500 A2 intelligent station, installing Ascend Docker Runtime requires modifying the Docker configuration. Run the vi /etc/sysconfig/docker command to delete the --config-file="" parameter, and run systemctl restart docker to make the configuration take effect.

• The MEF service pre-installed on the Atlas 500 A2 intelligent station performs security hardening on Docker. Ascend Docker Runtime does not support running in a security-hardened Docker environment. If you need to use Ascend Docker Runtime, manually uninstall the MEF service by referring to the "Uninstalling MEF Edge" in the MindEdge Framework User Guide.

  Run the systemctl status docker command. If the returned information contains the /docker_entrypoint.sh field, Docker has been security-hardened by the MEF service.

Other Cluster Scheduling Components

The dependencies for the Arm architecture and x86_64 architecture are different. The cluster scheduling components support both IPv4 and IPv6, and use IPv4 by default.

Table 1 Software environment

./Ascend-hdk-<chip_type>-npu-driver_<version>_linux-<arch>.run --full --install-for-all

NOTE

• Choose to install Docker or Containerd based on the actual service scenario.
• For installing the operating system on Atlas servers, see the Installation Guide (Arm) and Installation Guide (x86_64). The installation guides do not cover all the operating systems mentioned above and are for reference only.
• Atlas A2 training series products have different operating system requirements in virtual machine scenarios. For specific operating system constraints, see the "Installation and Uninstallation on Virtual Machines" section in the Atlas A2 Center Inference and Training Hardware 26.0.RC1 NPU Driver and Firmware Installation Guide.

Networking Requirements

Since Volcano, the core scheduling component for cluster scheduling, is currently deployed on the management node of K8s, the following recommendations are made for deploying the management node based on K8s deployment requirements to ensure service health and stability. Customers can make adjustments based on their own service characteristics.

• Management nodes should be separated from compute nodes and storage nodes. It is recommended to use a dedicated server for deployment.
• If the cluster is large in scale or has high requirements for service reliability, multiple management nodes are required.

Deployment Logic Diagram

Figure 1 Deployment logic diagram

The node types in a data center cluster are generally classified into the following three types:

• Management node (Master node): Manages the cluster and is responsible for distributing training and inference jobs to each compute node for execution. Cluster scheduling components associated with the master node can be installed.
• Compute node (Worker node): Executes training and inference jobs. Cluster scheduling components associated with the worker node can be installed.
• Storage node: stores data such as datasets and models output from training.

Users need to divide the network planes into:

• Service plane: Used for K8s cluster service management.
• Storage plane: Used to read training datasets from storage nodes. Because there are certain bandwidth requirements, it is recommended to use a separate network plane and network port to connect training nodes (management nodes or compute nodes) with storage nodes.
• Parameter plane: Used for parameter exchange between training nodes during distributed training. Refer to the following networking instructions.
  • Ascend Training Solution Networking Guide: Provides instructions on building networks for Huawei training computing devices (including Atlas 800 training servers, Atlas 900 PoD (model 9000), etc.).
  • Ascend Training Solution Networking Guide (Atlas A2 Training Products): Provides instructions on building networks for Huawei training computing devices (including Atlas 800T A2 training servers, Atlas 900 A2 PoD cluster basic units, and training servers integrated with the Atlas 200T A2 Box16 heterogeneous subrack).

Software and Hardware Specifications

Operating System Disk Partition

The recommended operating system disk partition is shown in Table 2.

Table 2 Disk space planning

Hardware Specifications

The hardware must meet the following requirements:

Table 2 Resource requirements

Cluster Scheduling Component Resource Configuration Requirements

The cluster scheduling component resource configuration must meet the following requirements:

Table 3 Resource configuration requirements for components installed on management nodes

Table 4 Resource configuration requirements for components installed on compute nodes