High-Frequency Telemetry

High-Frequency Telemetry (HFT) samples switch counters at intervals as short as 100 microseconds, on demand. Regular telemetry arrives every 10 or 30 seconds, which hides the sub-second bursts behind problems such as slow BGP convergence, multiplane imbalance, and NCCL stragglers. With HFT, you start a collection on the switches, ports, and counters you choose, for a set length of time. NVIDIA Minerva configures the switches, stores every sample, and returns each switch to its previous configuration when the collection ends.

How HFT Works

1. You start a collection from the UI or the HFT API.

2. NVIDIA Minerva sends the HFT configuration to each switch through the switch's collector, over the switch's NVUE REST API.

3. Each switch streams its samples to the collector, which forwards them to the platform.

4. The platform stores every raw sample, with no downsampling, and writes 10-second minimum and maximum summaries to a separate HFT data source in Trends.

5. When the collection ends, NVIDIA Minerva removes the HFT configuration from each switch.

HFT data has its own ingestion path, separate from regular telemetry. A collection doesn't change or slow regular telemetry from the same switch.

Requirements

• NVIDIA Spectrum-4 and later switches running Cumulus Linux 5.18 or later.

• An on-premises Docker Swarm installation of NVIDIA Minerva: either a single-node installation, or a platform with remote collectors. HFT is installed and enabled automatically. Kubernetes installations do not support HFT.

HFT runs without PTP, but to keep each switch's clocks within 0.5 seconds of the correct time, configure PTP as Cumulus Linux requires. See Prepare Switches for HFT.

Limits

• Sampling interval: 100 microseconds to 12.75 milliseconds, in steps of 50 microseconds. In the UI, other values are rounded up to the next 50 microseconds. The HFT API refuses them.

• Duration: 1 second to 1 hour. The collection stops automatically at the end.

• Counters: up to 20 per collection.

• One collection per switch: a switch can run only one HFT collection at a time. There is no fixed limit on the number of switches.

• Data rate: the platform accepts a limited number of data points per second across all active collections (4,000,000 by default) and per switch (200,000 by default). A collection that would go over either limit is refused. See Set HFT Capacity Limits.

• Retention: raw samples are kept for 12 hours by default (you set this during installation), and the 10-second summaries for 10 days. Raw samples are removed by age. HFT data on the platform is capped at 1 TB of disk; if the cap is reached first, older samples are removed sooner.

Permissions

HFT is available in user tenants as well as the admin tenant.

• Administrators and users with write access can start collections.

• The user who started a collection can stop it. A Platform Admin, or an Admin of that user's tenant, can also stop it.

• Read-only users can't open collections. They can still explore HFT data: in Trends, they change the data source to the HFT data source.

• Only Platform Admins can change capacity limits and manage switch credentials.

Set Up HFT

1. Prepare each switch. See Prepare Switches for HFT.

2. Register the switch credentials on the platform. See Register Switch Credentials for HFT.

3. Start a collection. See Run an HFT Collection.

NVIDIA Minerva configures HFT on the switches for you. For details about HFT on the switch itself, see High Frequency Telemetry in the Cumulus Linux 5.18 documentation.

HFT relies on each collector having its own collector ID. Collectors installed from a named bootstrap bundle already have one. See Install Collectors for a Hybrid Deployment.

Last updated: October 30, 2026