Cloud Computing · SS2026

Edge-Computing Monitoring

An edge-computing system that detects people and threats — fire, smoke — at the edge on a Raspberry Pi cluster, and surfaces events through a live web frontend and Telegram notifications. This site documents all ten seminar tasks: what was built and how it works.

9-node Raspberry Pi cluster k3s / Kubernetes YOLOv8n fire/person/smoke MPI · HPL · Amdahl & Gustafson Prometheus + Grafana MinIO · PostgreSQL · MQTT

What the system does

Compute is split across two layers. Hardware nodes — a camera-equipped Raspberry Pi and a Hailo AI accelerator — run as host services directly on the Pis, because they're bound to physical devices. Everything else runs as a Kubernetes (k3s) workload: the backend API, object storage, database, MQTT broker, frontend, and the Telegram notifier. See System Architecture for the full breakdown and the end-to-end event flow.

The ten tasks

Ten tasks from the course brief, each mapped to where it lives in this repository.

TASK 1

Infrastructure

PXE-diskless Raspberry Pi cluster: Pi 5 control plane, eight Pi 3 workers netbooting over TFTP/NFS.

TASK 2

Performance Benchmarking

HPL (LINPACK) GFLOPS measurements across 1, 2, 4, and 9-node configurations.

TASK 3

MPI Deployment & Scalability

OpenMPI across the Pi 3 workers; Amdahl and Gustafson demonstrated with two custom MPI apps.

TASK 4

Scaling Laws (non-MPI)

POV-Ray row-split rendering harness demonstrating Amdahl's and Gustafson's Laws without MPI.

TASK 5

System Monitoring

Prometheus + Grafana, Node Exporter on every Pi, ten alert rules piped to Telegram.

TASK 6

Object Detection Model

Custom-trained YOLOv8n on a 30k-image fire/person/smoke dataset — not a pretrained model.

TASK 7

Backend

FastAPI on k3s, 3 replicas, replicated PostgreSQL (CloudNativePG) + distributed MinIO.

TASK 8

Frontend

React app with a live, MQTT-pushed event log of detections and evidence images.

TASK 9

Telegram Notifications

Telegram bot service with subscriber self-registration and dry-run mode for testing.

TASK 10

Documentation & Risk Register

This site, plus an unusually candid honesty document about what the HA design does and doesn't survive.

Repository layout

Path	What	Task
`PXE-Boot/`	Diskless PXE network-boot setup for the Pi 3 workers	1
`HPL/`	High-Performance LINPACK GFLOPS benchmark	2
`MPI/task_3/`	MPI benchmark tutorial + Amdahl/Gustafson results	3
`PovRay/`	Non-MPI scaling-law demo (POV-Ray render split)	4
`MonitoringPIs/`	Prometheus + Grafana monitoring stack	5
`ai-model/`	YOLOv8n fire / person / smoke model: training + detection	6
`backend/`, `cluster/`	FastAPI backend, k3s manifests, MinIO + PostgreSQL	7
`frontend/`	React/Vite web UI	8
`telegram-bot/`	Telegram notifier service	9
`docs/`, `pages/`	Architecture docs, risk register, this site	10
`sensor-node/`, `inference-node/`	Camera streamer + Hailo inference — host services on the Pis	1, 6

Where this content comes from

Every page here is written from the module README.md files and the design docs already in this repository (docs/architecture.md, docs/storage.md, docs/flow.md, docs/weaknesses.md, plus each subtask's own README).