Azure OpenAI Realtime (WebRTC) Load Test

Minimal Node.js project to measure handshake performance and reliability when establishing many concurrent Azure OpenAI Realtime (WebRTC) data channel sessions.

UI (gpt-4o-realtime-preview RPM: 36)

flow overview

|

What It Does

• Repeatedly creates WebRTC peer connections against the Azure Realtime endpoint.
• Measures time from start until DataChannel open (handshake time).
• Aggregates metrics (success / failure / p95 / last 10) and streams them to a small browser dashboard via Server‑Sent Events (SSE).
• Lets you ramp up connections in batches (ramp step + interval + per‑connection spacing + optional jitter).

Architecture

Browser (dashboard) <--SSE / HTTP--> Express Server --(HTTP + WebRTC signaling)--> Azure OpenAI Realtime

Key files:

• server.js Express API + ramp logic
• webrtcSession.js Creates one WebRTC session (ephemeral key -> SDP offer/answer -> DataChannel)
• metrics.js Simple aggregator
• public/index.html Dashboard UI

Endpoints

Method	Path	Description
POST	/start	Begin a run (JSON body with parameters)
POST	/stop	Stop and close all active connections
GET	/events	SSE stream of metrics every second
GET	/status	One‑shot metrics JSON
GET	/	Dashboard page

Run Parameters (POST /start)

{
  "target": 50,            // total desired concurrent web socket connections
  "rampStep": 20,           // connections per batch
  "rampIntervalMs": 5000,   // wait between batches
  "batchDelayMs": 200       // delay between spawns inside one batch
}

Server side validation enforces: rampStep >= 1, rampIntervalMs >= 5000, batchDelayMs >= 200.

Environment Variables (.env)

See .env.example for a template you can copy. Do NOT commit your real .env.

Name	Required	Notes
AZURE_OPENAI_RESOURCE	yes	Resource name (without https://)
AZURE_OPENAI_KEY	yes	Resource API key (used to mint ephemeral key)
AZURE_OPENAI_DEPLOYMENT	yes	Deployment logical name (model)
AZURE_OPENAI_REGION	yes	Region for Realtime endpoint (e.g. swedencentral)
API_VERSION	yes	API version supporting realtime preview
DC_TIMEOUT_MS	no	DataChannel open timeout (default 8000)
VERBOSE_LOG	no	"true" for extra console logging
LAUNCH_JITTER_MS	no	Adds random 0..N ms before each spawn
PORT	no	HTTP port (default 3000)

Example .env (do NOT commit):

AZURE_OPENAI_RESOURCE=myresource
AZURE_OPENAI_KEY=***
AZURE_OPENAI_DEPLOYMENT=gpt-realtime
AZURE_OPENAI_REGION=swedencentral
API_VERSION=2025-04-01-preview
DC_TIMEOUT_MS=8000
VERBOSE_LOG=false
LAUNCH_JITTER_MS=50
PORT=3000

Quick Start

1. Clone repo and create .env from the example above (or copy .env.example).
2. Install deps:

   npm install

3. Run server:

   node server.js

4. Open http://localhost:3000 in a browser, adjust parameters, press Start.
5. Watch metrics update live. Press Stop to terminate all connections.

Metrics Explained

• started: attempted connection count
• success: DataChannel opened
• failed: any error before open
• active: in‑flight attempts (started - success - failed)
• avgHandshakeMs / p95HandshakeMs: aggregate latency
• last10: last 10 (sorted) successful handshake times
• errors: map of error message -> occurrences

Notes

• Adds a silent audio track because the service requires an audio m-line.
• Uses ephemeral key pattern: long lived key only used server side to request short‑lived key.
• TURN not required for direct connectivity in this test scenario.

Limitations / TODO

• No persistence of historical runs.
• No UI charting (only numbers) by default.
• No automatic backoff / retry logic.

License

MIT – see LICENSE.

---

Minimal, focused, and intentionally small for quick experimentation.