Skip to content

README.md

Latest commit

 

History

History
111 lines (89 loc) · 3.73 KB

README.md

File metadata and controls

111 lines (89 loc) · 3.73 KB

Request Speed Test

A high-throughput HTTP load testing project using the rnet library (Rust-based HTTP client with updated TLS settings to bypass WAF checks). This project demonstrates achieving over 20,000 requests per second (RPS) with proper system tuning and client configuration.

I was able to send 10 million requests in 8 minutes:

Overview

This project explores the limits of HTTP load testing with rnet, focusing on:

  • Overcoming client-side bottlenecks (file limits, ephemeral ports, connection reuse)
  • Server-side tuning for high concurrency (Nginx, kernel parameters)
  • Performance scaling across different server configurations
  • Error analysis and optimization

Key Findings

Client Limitations

  • rnet does not increase open file limits on Linux, causing failures with bursty requests
  • Requires manual tuning for high concurrency

Performance Results

  • Initial Results: ~1k RPS on local machine with default settings
  • After Tuning: 9k RPS on 4vCPU/8GB server
  • Peak Performance: 20k RPS on 32vCPU dedicated server
  • Stability: Maintains performance up to 10M+ requests

Server Configurations Tested

  • 2vCPU/4GB (shared): ~5k RPS
  • 4vCPU/8GB (shared): 9k RPS
  • 8vCPU (shared): 15k RPS
  • 32vCPU (dedicated): 20k RPS

Common Errors and Fixes

  • File Descriptor Limits: "Too many open files" → Increase ulimit and sysctl fs.file-max
  • Ephemeral Port Exhaustion: "Cannot assign requested address" → Expand net.ipv4.ip_local_port_range
  • HTTP/2 GOAWAY: Server closing connections → Increase worker_connections and http2_max_concurrent_streams
  • TLS Errors: Invalid certificate context → Disable verification for testing

Optimizations Applied

  • Client-Side:

    • Increased file descriptors to 65k
    • Expanded ephemeral ports to 64k range
    • Enabled TCP TIME_WAIT reuse
    • Adjusted client count for optimal multiplexing (1 client per 100 workers)

  • Server-Side:

    • Nginx worker_connections set to 65k
    • HTTP/2 streams increased to 1k per connection
    • Kernel somaxconn set to 65k
    • Systemd limits configured

Quickstart

Prerequisites

  • Linux server (Ubuntu 25.04 recommended)
  • Python 3.13+
  • Root access for system tuning

Setup

  1. Clone and Install:

    git clone https://github.com/lafftar/requestSpeedTest.git
cd requestSpeedTest
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install uvloop

  2. Tune Client System:

    sudo bash client/tune_server.sh
sudo reboot  # Required for limits to take effect

  3. Setup Server (if running on same machine):

    sudo bash server/startup_script.sh

Running Tests

source .venv/bin/activate
python -m send_request.rnet_test --total-requests 100000 --concurrency 10000

Example Outputs

  • 400k requests: 15k RPS
  • 1M requests: 8k RPS (stable)
  • 10M requests: Consistent performance

Project Structure

  • send_request/rnet_test.py: Main load testing script
  • client/: Client-side setup and tuning scripts
  • server/: Server-side Nginx configuration and tuning
  • utils/: Shared utilities (logging, limits)
  • notes.md: Detailed findings and test results with screenshots

Troubleshooting

  • If scripts fail, follow manual instructions in client/README.md and server/README.md
  • Ensure ulimit -n shows 65536 after reboot
  • Check server logs for Nginx errors
  • For high concurrency, monitor CPU and memory usage

Contributing

This project is focused on performance analysis. Contributions for further optimizations or additional HTTP clients welcome.

License

MIT License