# Goodput Monitor

> Goodput measures whether GPU hardware actually delivered the capability you paid for — clocks, ECC, InfiniBand, and Xid faults — and turns readings into availability weights and SLA claim packs. Hardware telemetry only; no workload data ever leaves the node.

Site: https://files-two-amber.vercel.app/
Current release: v0.9.0 (2026-09-14)
Docker images: ghcr.io/kjd07/goodput-agent:0.9.0, ghcr.io/kjd07/goodput-ingest:0.9.0

## Product

- [Landing page](https://files-two-amber.vercel.app/): Marketing site with install guides for Linux, Windows, and iOS dashboard viewing
- [Install guide](https://files-two-amber.vercel.app/docs/install.html): Styled install paths for Linux, Windows, Docker, and bare metal
- [Data contract](https://files-two-amber.vercel.app/docs/data-contract.html): What the agent transmits (hardware telemetry only)
- [Helm chart](https://files-two-amber.vercel.app/docs/helm.html): Kubernetes deployment guide
- [Full guide (plain text)](https://files-two-amber.vercel.app/llms-full.txt): Machine-readable copy for AI assistants

## Install (one-line)

- Linux GPU server: `curl -fsSL https://files-two-amber.vercel.app/install-linux.sh | bash`
- Windows (Docker Desktop): `irm https://files-two-amber.vercel.app/install-windows.ps1 | iex`
- iPhone/iPad: view-only — open the dashboard URL in Safari after someone runs the Linux or Windows installer on a host

## Downloads

- [install-linux.sh](https://files-two-amber.vercel.app/install-linux.sh): Full Docker stack + GPU agent on Linux
- [install-windows.ps1](https://files-two-amber.vercel.app/install-windows.ps1): Docker Desktop stack on Windows
- [docker-compose.release.yml](https://files-two-amber.vercel.app/docker-compose.release.yml): Pre-built compose file
- [releases/manifest.json](https://files-two-amber.vercel.app/releases/manifest.json): Version, artifacts, and platform metadata

## After install

Dashboard: http://localhost:8080/ (GPU data within ~60 seconds)
Default dev credentials: token `devtoken`, cluster `local-dev`

## Requirements

- Linux GPU nodes: Docker Engine, NVIDIA driver (`nvidia-smi`), NVIDIA Container Toolkit
- Windows: Docker Desktop with WSL 2, running before install
- Agent does not run on macOS or Windows; GPU monitoring stays on Linux nodes

## Privacy / data contract

The agent collects NVML GPU counters, InfiniBand port state/counters, and numeric Xid codes — never process names, command lines, env vars, filenames, container images, models, datasets, or memory contents.