Skip to content

Latest commit

 

History

History
137 lines (100 loc) · 4.88 KB

README.md

File metadata and controls

137 lines (100 loc) · 4.88 KB

ROS 2 Client Library Benchmarks

Performance benchmarks for comparing ROS 2 client libraries: C++ (rclcpp), Python (rclpy), and JavaScript (rclnodejs).

Prerequisites

  1. ROS 2: Install from ros.org
  2. Node.js: v16+ for rclnodejs (from nodejs.org)
  3. rclnodejs: Follow installation guide
  4. Build Dependencies: For C++ benchmarks: sudo apt install libssl-dev cmake build-essential

Benchmark Structure

Each client library has identical benchmark tests:

Test Type Description
topic Publisher/subscriber performance
service Client/service request-response performance

Performance Results

Test Environment

Hardware:

  • CPU: 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz (4 cores, 8 threads)
  • Memory: 8GB RAM
  • Architecture: x86_64

Software:

  • OS: Ubuntu 24.04.3 LTS (WSL2)
  • Kernel: 6.6.87.2-microsoft-standard-WSL2
  • ROS 2: Jazzy Patch Release 6
  • C++ Compiler: GCC 13.3.0 (Ubuntu 13.3.0-6ubuntu2~24.04)
  • Python: 3.12.3
  • Node.js: v22.18.0
  • rclnodejs: v1.5.0

Benchmark Results

Benchmark parameters: 1000 iterations, 1024KB message size

Client Library Topic (ms) Service (ms) Performance Ratio
rclcpp (C++) 168 627 Baseline (fastest)
rclpy (Python) 1,618 15,380 9.6x / 24.5x slower
rclnodejs (Node.js) 744 927 4.4x / 1.5x slower

Last updated: August 30, 2025

Notes:

  • Topic benchmarks: All libraries completed successfully with 1024KB messages
  • Service benchmarks: All libraries completed successfully with 1024KB responses
  • Performance ratios are relative to C++ baseline
  • C++ shows excellent performance as expected for a compiled language
  • Node.js performs significantly better than Python, likely due to V8 optimizations
  • Python service performance shows significant overhead with large payloads compared to topics

Running Benchmarks

C++ (rclcpp)

cd benchmark/rclcpp/
mkdir -p build && cd build
source ~/Download/ros2-linux/local_setup.bash  # Adjust path
cmake .. && make
# Run from build directory:
./publisher-stress-test -r 1000 -s 1024
./client-stress-test -r 1000

Python (rclpy)

cd benchmark/rclpy/
source ~/Download/ros2-linux/local_setup.bash  # Adjust path
python3 topic/publisher-stress-test.py -r 1000 -s 1024
python3 service/client-stress-test.py -r 1000

JavaScript (rclnodejs)

cd /path/to/rclnodejs/  # Project root
source ~/Download/ros2-linux/local_setup.bash  # Adjust path
node benchmark/rclnodejs/topic/publisher-stress-test.js -r 1000 -s 1024
node benchmark/rclnodejs/service/client-stress-test.js -r 1000

Test Workflow

For complete tests, run subscriber/service first, then publisher/client:

Topic Test:

# Terminal 1: Start subscriber (adjust for your language)
python3 topic/subscription-stress-test.py          # Python
./subscription-stress-test                          # C++ (from build dir)
node benchmark/rclnodejs/topic/subscription-stress-test.js  # Node.js

# Terminal 2: Run publisher benchmark
python3 topic/publisher-stress-test.py -r 1000 -s 1024     # Python
./publisher-stress-test -r 1000 -s 1024                    # C++ (from build dir)
node benchmark/rclnodejs/topic/publisher-stress-test.js -r 1000 -s 1024  # Node.js

Service Test:

# Terminal 1: Start service (adjust for your language)
python3 service/service-stress-test.py -s 1024             # Python
./service-stress-test -s 1024                              # C++ (from build dir)
node benchmark/rclnodejs/service/service-stress-test.js -s 1024  # Node.js

# Terminal 2: Run client benchmark
python3 service/client-stress-test.py -r 1000      # Python
./client-stress-test -r 1000                       # C++ (from build dir)
node benchmark/rclnodejs/service/client-stress-test.js -r 1000  # Node.js

Message Types

  • Topics: std_msgs/msg/UInt8MultiArray (configurable size)
  • Services: nav_msgs/srv/GetMap (map data request/response)

Notes

  • All benchmarks use high-precision timing for accurate measurements
  • C++ provides baseline performance; Python/JavaScript show expected interpreted language overhead
  • All implementations are now working correctly across the three client libraries
  • Results vary by system configuration; use relative comparisons between client libraries
  • Service benchmarks involve request-response cycles with substantial data payloads (OccupancyGrid maps)