Reference

Glossary

View as Markdown

Terms used throughout the RMS documentation.

BMC

Baseboard Management Controller - the out-of-band controller RMS reaches over Redfish for compute and power-shelf nodes.

CPLD .vme

Complex Programmable Logic Device firmware image format used for switch CPLD components, extracted by nvfwupd.

Firmware manifest

The JSON, distributed per firmware milestone, that describes board SKUs, components, versions, and artifact locations. RMS parses it into a firmware object. See the sample manifest.

Firmware object

Persisted workflow state representing a parsed firmware manifest for a hardware type: a software object tree, resolved components, cached artifact metadata, an optional default marker, and apply history.

.fwpkg

A PLDM firmware package containing one or more component images.

gNMI

gRPC Network Management Interface - used for switch telemetry/config connectivity checks.

HMC

Hardware Management Controller - a management endpoint on some NVIDIA trays, addressed as a firmware target alongside the BMC.

In-memory store

The default persistence backend when no database URL is configured. Fast and dependency-free, but all workflow state is lost on restart.

Insecure mode

Plaintext, unauthenticated gRPC. Requires both [tls] insecure = true and the RMS_ALLOW_INSECURE=1 env gate. Dev/test only.

Insecure switch

[switches] insecure_switch = true - RMS drops outbound switch client mTLS (unverified NVUE HTTPS, plaintext NMX-C). Dev/test only.

Job

An asynchronous unit of long-running work (firmware update, switch system-image install) tracked by ID. Clients poll job status until it reaches a terminal state.

JobTracker

The orchestrator component that owns job lifecycle: creation, status, retention (max_tracked_jobs), and TTL-based cleanup of terminal jobs (terminal_job_ttl_seconds).

librms

The external crate (nv-rms-client) providing the rack_manager.proto gRPC service definitions, pinned by rev and fetched by Cargo.

Multi-node NVLink topology data (chassis serial, slot, tray index) RMS reads from the Redfish Processor OEM block.

mTLS

Mutual TLS - the default for both the northbound gRPC API and outbound switch connections. Requires a server cert/key plus a client CA.

NMX-C

The NMX Controller / scale-up fabric manager on NVSwitch trays. RMS configures it and queries its state over gRPC (nmx_gateway_id identifies RMS on these requests).

Node

An ephemeral object representing one managed hardware component (a compute tray, NVSwitch tray, or power shelf). Nodes are created per request and hold their transport client and credentials; RMS stores no device state beyond an operation’s lifetime.

Node descriptor

The role / vendor / product_family attributes a client sends to select a node type at the gRPC boundary (e.g. compute / nvidia / gb200). See HCL: node descriptor translation.

Node type

The concrete hardware identity RMS dispatches on, e.g. compute_gb200_nvidia, switch_gb200_nvidia, powershelf_gb200_delta. See the HCL.

nvfwupd

The embedded firmware-update engine (library + CLI). RMS drives it through its workflow API; the CLI also ships in the release image for diagnostics.

NVUE / NVOS

NVIDIA’s switch REST API (NVUE) and operating system (NVOS) on NVSwitch trays. RMS talks to switches over NVUE REST plus SSH/SFTP.

Parent/child job

A batch operation creates its parent job first, then attaches one child job per admitted node, so callers can track the batch as a whole and per node.

Persistence backend

The swappable store behind the FirmwareObjectStore trait: an in-memory store (default, volatile) or Postgres (durable).

PLDM

Platform Level Data Model - the firmware-package format (.fwpkg) parsed by the embedded nvfwupd crate.

Product family

The rack generation a node belongs to: GB200, GB300, or VRNVL72. RMS rejects nodes whose product family doesn’t match the rack type.

Rack

A trait-based container that owns a map of nodes and a node factory. RMS ships NvlGb200Rack, NvlGb300Rack, and NvlVrnvl72Rack.

Redfish

The DMTF standard HTTP/HTTPS management API RMS uses to talk to compute and power-shelf BMCs.

RMS

Rack Management Service - the stateless, rack-level hardware management service documented here. Exposes a single RackManager gRPC API.

Scale-up fabric manager

The NMX-C-managed NVLink fabric. RMS can query cluster state, enable/disable it on switch targets, and control telemetry interfaces.