Linux hosts

GPU monitoring for the boxes doing the expensive work.

On hosts with NVIDIA GPUs, the collector reads utilization, memory, and temperature per GPU through nvidia-smi and applies warn/fail thresholds: the same issue pipeline as every other check, on the hardware that costs the most.

What RootTrace watches

Every reading is checked continuously; bad readings become grouped issues with the evidence attached. RootTrace also remembers whether it has seen the same incident before.

Utilization per GPUPercentage busy, indexed per GPU on multi-GPU hosts.
Memory pressureGPU memory used as a percentage, with warn/fail thresholds.
TemperatureDegrees Celsius per GPU against a configurable fail threshold.
GPU inventoryGPU count and per-index identity, so a vanished device is an issue.

Warn and fail thresholds, prefilled

Sensible defaults out of the box, overridable per host with environment variables. No threshold spreadsheet required on day one.

Memory warn / fail percent ROOTTRACE_NVIDIA_MEMORY_WARN_PERCENT
Temperature fail °C ROOTTRACE_NVIDIA_TEMP_FAIL_C

Setup

nvidia gpu setup

# automatic where nvidia-smi is present; no flags needed.
# tune thresholds if your cards run hot by design:
$ export ROOTTRACE_NVIDIA_TEMP_FAIL_C='90'

Read-only by architecture: no write path, no remediation executor, no inbound ports. Credential-like evidence is redacted before it leaves the host.

One collector, live in minutes

Everything on this page comes from the same read-only collector: one command per host, then point it at your targets.

install on any Linux host

$ # after adding the signed RootTrace dnf or apt repository:
sudo dnf install roottrace-collector  # apt: sudo apt install roottrace-collector
sudoedit /etc/roottrace/collector.env  # set ROOTTRACE_COLLECTOR_TOKEN and API URL
sudo roottrace-collector-setup apply
sudo systemctl enable --now roottrace-collector

Questions, answered honestly

Does it need CUDA or a driver hook?

No. It shells out to nvidia-smi, the tool already on any host with the NVIDIA driver installed. If nvidia-smi is absent the check stays quiet.

Is this safe to run next to production?

The collector is read-only by architecture, not by policy: there is no write path, no remediation executor, and no privileged mode to switch on. It sends short-lived outbound HTTPS requests, opens no inbound ports, and redacts credential-like evidence before anything leaves the host.

Do I need a separate agent for this?

No. One RootTrace collector per host covers every integration on this site: databases, web servers, containers, endpoints, and the Linux host itself, all from one signed package or container. You point it at targets with environment variables; there is nothing else to deploy.

Free on 5 hosts. Live in minutes.

Create a workspace, install one collector, and watch issues explain themselves.

Create free workspace Schedule a demo