Quick diagnostics

更新时间:
复制 MD 格式

Quick Diagnostics lets you check the network and hardware health of Lingjun nodes in PAI Lingjun Intelligent Computing Service, and run network performance tests using multiple communication libraries and communication models.

What's covered

Diagnostic typeChecksWhen to use
Static configuration checkSystem software, network, GPU configurationBefore starting a training job, or when onboarding a new node
Dynamic runtime checkTCP connectivity, TCP latency, Remote Direct Memory Access (RDMA) connectivityWhen you suspect a live network issue during or after a job
System hardware diagnosticsHardware status of Lingjun nodesWhen you suspect a hardware fault
General network testRDMA traffic and latency between nodesTo baseline or benchmark cluster network performance
Communication library testCollective communication performance using ACCL (Alibaba Cloud Collective Communication Library) or NCCL (NVIDIA Collective Communications Library)To verify collective communication health before distributed training
Network diagnostics applies to instances in billing state only.

Self-service diagnostics

Network diagnostics

Static configuration check

Runs static configuration diagnostics on Lingjun nodes, covering system software, network, and GPU checks.

  1. Log on to the Lingjun console.

  2. In the left navigation pane, choose Quick Diagnostics > Self-service Diagnostics.

  3. Click the Network Diagnostics tab.

  4. Click Static Configuration Check.

  5. In the Diagnostic Information area, select the target Cluster Name and Node ID from the drop-down lists.

  6. Click Start Diagnostics.

Dynamic runtime check

Runs dynamic runtime diagnostics on Lingjun nodes, covering TCP connectivity, TCP latency, and RDMA connectivity.

  1. In the left navigation pane, choose Quick Diagnostics > Self-service Diagnostics.

  2. Click the Network Diagnostics tab.

  3. Click Dynamic Runtime Check.

  4. In the Diagnostic Information area, select the target Cluster Name and Node ID from the drop-down lists.

  5. Click Start Diagnostics.

Server diagnostics

Checks the hardware status of Lingjun nodes and displays results in a visual format.

  1. In the left navigation pane, choose Quick Diagnostics > Self-service Diagnostics.

  2. Click the Server Diagnostics tab.

  3. Click System Hardware Diagnostics.

  4. In the Diagnostic Information area, select the target Cluster Name and Node ID from the drop-down lists.

  5. Click Start Diagnostics.

Network tests

General network test

Measures RDMA traffic and latency between Lingjun nodes.

  1. In the left navigation pane, choose Quick Diagnostics > Network Test.

  2. Click the General Network Test tab.

  3. In the Test Information area, select a Network Protocol and Test Type, then configure the parameters.

  4. Click Start Diagnostics.

RDMA traffic test parameters

ParameterDescription
Traffic ModelMtoN model: Tests one-way connectivity from Clients nodes to Servers nodes, including single-to-single and many-to-many configurations. Fullmesh model: Tests connectivity between every pair of target Lingjun nodes.
Test DurationFixed duration selected from the drop-down list. Unit: seconds.
QPNumber of test process streams. Affects test bandwidth.
GDRWhen enabled, attaches Lingjun network interface controllers (NICs) to the corresponding GPUs for the traffic test.
Cluster NameThe cluster where the target Lingjun nodes are located.
Clients(MtoN model) Select the Clients nodes.
Servers(MtoN model) Select the Servers nodes.
Select Target Nodes(Fullmesh model) Select the target nodes.
Node PortThe starting port used for the test.

RDMA latency test parameters

ParameterDescription
Cluster NameThe cluster where the target Lingjun nodes are located.
Test NodesChecks network latency between every pair of the selected nodes.
Node PortThe starting port used for the test.

Communication library test

Tests collective communication performance using ACCL or NCCL across target Lingjun nodes.

  1. In the left navigation pane, choose Quick Diagnostics > Network Test.

  2. Click the Communication Library Test tab.

  3. Configure the Test Information.

  4. Click Start Test.

Communication library test parameters

ParameterDescription
Communication Library CategoryACCL (Alibaba Cloud Collective Communication Library) or NCCL (NVIDIA Collective Communications Library).
Communication ModelThe collective communication operation to test. See the table below for details.
Number of GPUsThe value range is 1 to 8.
Cluster NameThe cluster where the target Lingjun nodes are located.
Select Target NodesSpecify the IP address of each target node. No need to specify the Lingjun node to start the test.
Node PortThe starting port used for the test.

Communication models

ModelWhat it doesTypical use case
AllReduceAggregates data from all processes, reduces it to a single value, and distributes the result back to all processes.Gradient synchronization — the most common operation in distributed training.
AllGatherGathers data from all processes into a single structure accessible by every process.Collecting distributed shards (e.g., model parameters, embeddings).
AllGatherAExtends AllGather with data type parameters, supporting large and custom data types.Use when AllGather doesn't cover your data type requirements.
AllToAllEach process distributes its data to all other processes, and receives data from all others.Data redistribution across nodes, such as expert routing in MoE models.
AllToAllAExtends AllToAll with data type and buffer parameters, supporting data of different sizes and types.Use when communicating variable-size data across processes.
BroadcastDistributes data from one process to all other processes.Parameter initialization or broadcasting control signals.

View diagnostic reports

The diagnostic history shows reports for all Self-service Diagnostics and Network Tests. Each report includes a Report ID and Cluster Name. Click the tabs to filter reports by diagnostic type.

For any report, the Operations column provides the following actions:

  • View Report: View the full results and details of the diagnostic run.

  • Diagnose Again: Re-run the same diagnostic.