GAPS: GPU-Aware Pub/Sub Communication

This repo contains the framework introduced in the paper "Towards Low-Latency GPU-Aware Pub/Sub Communication for Real-Time Edge Computing" accepted by RTCSA 2025.

Implementations

There are two implementations of GAPS in different branches:

• main: GAPS-z, built on top of Zenoh-cpp (with Zenoh-pico as backend)
• iceoryx: GAPS-i, built on top of Iceoryx.

Environment

Docker environment is priovded for development or testing. To set it up, make sure the following requirements are met:

• Docker is installed
• Docker Compose is installed
• NVIDIA Container Toolkit is installed
  • Follow the installation and configuration steps here
• CUDA 12.6 is supported by your host's NVIDIA driver
• CUDA Driver APIs like cuMemCreate and cuMemExportToShareableHandle are supported by your NVIDIA GPU

There are two environments under env/:

• x86: for x86 machines with an NVIDIA GPU
• jetson: for NVIDIA Jetson embedded system like Jetson AGX Orin
  • ⚠️ Please see env/jetson/README.md for how to build its base image before proceeding

Change directory to either one, and then run the following commands:

docker compose up -d
ssh ubuntu@localhost -p 22222

To destroy the environment, run the following command under the same directory:

docker compose down

Build

To configure build instructions with CMake and build the project with Ninja, run the following commands in the parent directory of GAPS/:

cmake GAPS -B build -G Ninja
ninja -C build

The compiled executables will be generated under ./build/src.

CMake Build Options:

• PROFILING=[on|(off)]: whether to profile publisher's put and subscriber's callback
• BUILD_DEBUG=[on|(off)]: whether to build with debugging codes
• BUILD_TORCH_SUPPORT=[on|(off)]: whether to build PyTorch support (i.e., to build PyGAPS)
• BUILD_EXAMPLES=[(on)|off]: whether to build example codes

Note

The one wrapped with parentheses is the default value

Pre-Commit

Pre-commit is used to setup clang-format pre-commit hooks.

1. Install pre-commit in your Python virtual environment.
2. Run pre-commit install to install the hooks.
3. Each time before committing, committed files will be formatted with clang-format.

Acknowledgements

Thanks for the following works to make this project possible.

• This project depends on the following third-party libraries:
  • CUDA
  • Zenoh-cpp
  • Zenoh-pico
  • Iceoryx
  • nanobind
• Example codes in this projects depend on the following third-party libraries:
  • OpenCV
  • PyTorch
  • ultralytics
• Thanks tlsf-bsd for showing how to implement the TLSF allocator.
• Thanks jetson-containers for providing machine learning containers on NVIDIA Jetson embedded systems.