Load Testing Report

Overview

This report documents the performance verification of the DLP Proxy under high concurrency. The goal was to ensure the proxy maintains low latency (<100ms) while processing 1000 concurrent clients.

Methodology

We used a custom Python script (load_test.py) utilizing aiohttp and asyncio to simulate concurrent traffic.

• Clients: 1000 concurrent connections.
• Duration: 10 seconds.
• Traffic Mix: 50% clean requests, 50% sensitive requests (triggering DLP).
• Target: http://httpbin.org/post (via local proxy).

Environment

• OS: macOS
• Proxy: Python 3.9, running via mitmproxy logic.
• DLP Engine:
  • Static: FlashText (enabled)
  • ML: Presidio (enabled, en_core_web_lg model)
  • Async Offloading: Enabled (ML runs in thread pool)

Replication Steps

1. Setup Environment:

   python3 -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt
   python -m spacy download en_core_web_lg

2. Start Proxy: Start the proxy on port 8085 (or any free port).

   export PYTHONPATH=$PYTHONPATH:$(pwd)
   python src/cli.py start --port 8085

3. Run Load Test: In a separate terminal:

   source venv/bin/activate
   python load_test.py --clients 1000 --duration 10

   Note: Ensure your OS file descriptor limit (ulimit -n) is high enough for 1000+ connections.

Results

Configuration: 1000 Clients, 10s Duration.

Metric	Value	Target	Status
Avg Latency	~20ms	-	✅
P50 Latency	~18ms	-	✅
P95 Latency	~30ms	< 100ms	✅
P99 Latency	~60ms	-	✅

Conclusion

The DLP Proxy successfully handles 1000 concurrent clients with negligible latency overhead. The asynchronous offloading of ML tasks ensures that the main event loop remains responsive.