This guide covers deploying the NVIDIA Minerva platform with Docker Swarm, in either the unified or hybrid deployment model. To install on a Kubernetes cluster instead, see Kubernetes Installation.
Collectors connect to the platform over a management websocket on port 443. The platform doesn't authenticate clients on this websocket, which carries collector registration, re-registration, and the configuration the platform sends to collectors. Other management APIs require a UI session or an API key. Restrict access to the platform's HTTP and HTTPS ports (80 and 443) to trusted networks, including the collector hosts, with your firewall or security-group rules.
Unified Installation
In a unified deployment, the NVIDIA Minerva stack runs both the platform and collector roles in a single installation. The following ports must be open on unified installations:
|
Port |
Protocol |
Purpose |
|---|---|---|
|
443 |
TCP |
UI and API (HTTPS) |
|
80 |
TCP |
UI and API (HTTP) |
|
4317 |
TCP |
OTLP gRPC (device telemetry) |
|
4318 |
TCP |
OTLP HTTP (device telemetry) |
|
4319 |
TCP |
OTLP gRPC (NMX-T; NVLink ingress) |
|
4320 |
TCP |
DMT OTLP gRPC ingress (optional) |
|
4321 |
TCP |
OTLP gRPC (high-frequency telemetry) |
|
9091 |
TCP |
Prometheus remote-write (HTTP push, not a scrape endpoint) |
|
57402 |
TCP |
gNMI dial-out (WJH) |
|
514 |
UDP |
Syslog |
For high-frequency telemetry (HFT), the host must also be able to reach the NVUE REST API on each Cumulus switch, on TCP 8765 or the port in the switch's HFT credentials.
On a single-node unified installation, the Docker Swarm ports (2377, 7946, and 4789) are bound but have no peers, so they don't need to be open in the firewall.
Hybrid Installation
In a hybrid deployment, the NVIDIA Minerva Platform Stack and Collector Stack are deployed separately. The Platform Stack can run either on-premises or in the cloud, while the Collector Stack typically runs on-premises. The Collector Stack sends telemetry data to the Platform Stack over a local network or the internet. Each component is installed independently, and you select the installation type at an interactive prompt.
Note: Cloud and SaaS deployments of NVIDIA Minerva are not currently supported.
Components:
-
Platform Stack: the platform services, deployed on on-premises hosts or in the cloud
-
Collector Stack: collectors and ingestion services deployed on on-premises hosts, configured to reach the Platform Stack
The following ports must be open for hybrid installations.
Inbound to Collector Nodes (from Devices)
|
Port |
Protocol |
Purpose |
|---|---|---|
|
4317 |
TCP |
OTLP gRPC (device telemetry) |
|
4318 |
TCP |
OTLP HTTP (device telemetry) |
|
4319 |
TCP |
OTLP gRPC (NMX-T; NVLink ingress) |
|
4321 |
TCP |
OTLP gRPC (high-frequency telemetry) |
|
9091 |
TCP |
Prometheus remote-write (HTTP push, not a scrape endpoint) |
|
57402 |
TCP |
gNMI dial-out (WJH) |
|
514 |
UDP |
Syslog |
Inbound to Platform Nodes (from Collectors)
|
Port |
Protocol |
Purpose |
|---|---|---|
|
443 |
TCP |
UI, API, management (HTTPS), including the management websocket that collectors register on |
|
9093 |
TCP |
Kafka (mTLS). Reached at the Kafka address, which can differ from the platform address |
|
14317 |
TCP |
Data-plane metrics (mTLS) |
|
14321 |
TCP |
High-frequency telemetry (mTLS) |
|
24317 |
TCP |
Collector health metrics/logs (mTLS) |
Run only one installer at a time on each host. Running installers in parallel on the same host can make the installation fail.
Last updated: