Skip to content

Cnuas Management Tools, Datasheet

Item Value
Part cnuassmi, cnuas-cli, cnuas-link-cli, cnuas-webui, cnuas-rackmon
Type Operator clients
Packages cnuas-gpu-modules, cnuas-cli, cnuas-link-cli
Version fd0f896-dirty
Repos PacketFive/CnuasGPU, PacketFive/CnuasSwitch, PacketFive/CnuasLink, PacketFive/cnuas

1. Overview

These are the small, focused clients that sit next to each component. They exist because an operator debugging one daemon should not have to go through the whole control plane, and because the equivalents on real hardware are the tools people already know. The unified surface across all components is the control plane; everything here is the local alternative.

Key features

  • Accelerator inventory and telemetry in the familiar tabular form.
  • Machine readable output for scripting, in comma separated and JSON forms.
  • Repeating sampling mode for watching a value change.
  • Direct clients for the fabric switch and the accelerator link switch.
  • Rack power shelf client for the management controller.

2. Tool summary

Tool Talks to Transport
cnuassmi Accelerators libcnuasrt and the accelerator sysfs class
cnuas-cli Fabric switch daemon JSON over UNIX domain socket
cnuas-link-cli Accelerator link switch daemon JSON over UNIX domain socket
cnuas-webui Fabric switch daemon Browser dashboard over the management socket
cnuas-rackmon Rack management daemon JSON over UNIX domain socket

3. Accelerator management interface, cnuassmi

The equivalent of nvidia-smi for CnuasGPU. Called with no arguments it prints a table covering every visible accelerator: identity, revision, memory in use against memory installed, multiprocessor and lane counts, and fabric link state.

Option Effect
--query-gpu Select named fields instead of the table
--format Choose csv or json output
--noheader Omit the header row
--loop=N Repeat every N seconds until interrupted
--version Report tool, driver and runtime versions
--help Usage

Queryable fields

index, name, gpu_id, revision, fw_version, sm_count, lanes_per_sm, tensor_size, memory.total, memory.used, memory.free, link.up, link.rx_ready.

Live memory usage is read from the driver sysfs attributes rather than through an extra device call, so sampling in a loop stays cheap.

4. Fabric switch client, cnuas-cli

Drives the CnuasSwitch management socket directly: port status and mode, link state, the Ethernet forwarding database, the InfiniBand linear forwarding table, data centre bridging settings, the subnet manager, and telemetry.

Drives the CnuasLink switch daemon: port state, the accelerator forwarding database, link state, telemetry and version.

6. Dashboard, cnuas-webui

A browser dashboard over the fabric switch, with a backend that speaks the management socket and a frontend that renders port state and counters. It is a laboratory tool and carries no authentication of its own.

7. Rack power client, cnuas-rackmon

Queries cnuas-rackmond over its UNIX domain socket and formats the power supply and battery backup tables for the shelf. See CnuasBMC.

8. Planned tools

These are named in the design documents and have no implementation.

Tool Models Blocked on
cnuas-dcgmi, cnuas-dcgm Data centre GPU manager Telemetry schema
cnuas-prof Kernel level profiler Performance counter interrupt path and CnuasCC

9. Validation

Tool Coverage
cnuas-cli Exercised through the switch management interface tests, tests/test_mgmt_api.py, 21 cases
cnuas-link-cli src/cnuaslink/cli/tests/test_cli_e2e.py, 10 cases
cnuassmi No automated test, requires an accelerator device
cnuas-webui No automated test
cnuas-rackmon No automated test

See the Validation Matrix.

10. Integration information

Item Value
cnuassmi source src/cnuasgpu/tools/cnuassmi/, with a manual page
cnuas-cli source src/cnuasswitch/cli/
cnuas-link-cli source src/cnuaslink/cli/
cnuas-webui source src/cnuasswitch/webui/
cnuas-rackmon source bmc/rmc/src/
Language C, with a web frontend for the dashboard

11. Revision history

Revision Notes
A First publication. Options and queryable fields read from the tool sources.