MPI + CUDA + OpenMP Professional Template

A high-performance computing template designed for distributed GPU workloads using MPI (Message Passing Interface), CUDA, and OpenMP. This template provides a robust build system, optimized memory management, and easy multi-node configuration.

Features

• Modern CMake: Target-based configuration with C++17, CUDA 11+, and OpenMP support.
• Hybrid Parallelism: Seamless integration of MPI (inter-node), OpenMP (intra-node CPU), and CUDA (intra-node GPU).
• Environment Configuration: Robust workload and build management via .env files.
• Optimized Memory: Built-in PinnedAllocator for high-performance H2D/D2H transfers.
• Monitoring Tools: monitor.sh for multi-node GPU monitoring using tmux and ssh.
• Utility Suite: Colorized logging, type-safe environment parsing, and CUDA error handling.

Directory Structure

• src/: Core implementation logic (.cpp, .cu).
• include/: Public headers.
• utils/: Reusable HPC utilities (allocators, macros, etc.).
• examples/: Main application entry points.
• lib/: Third-party submodules (e.g., Google Benchmark).
• build/: Generated build artifacts.

Quick Start

1. Configure

Create your local environment file from the example:

cp .env.example .env

2. Build

make build

3. Run

make run

Environment Variables (.env)

The template automatically loads variables from .env. These affect both compilation and runtime.

Advanced Usage

GPU Architecture

The Makefile and CMakeLists.txt use the -real suffix (e.g., 75-real) to prevent generating PTX code, which avoids "unsupported toolchain" errors on systems with older drivers while maintaining maximum performance.

Multi-node Monitoring

Use the included monitor.sh to watch GPU status across your cluster:

./monitor.sh "watch -n 0.1 nvidia-smi"

Note: Update the HOSTS array in monitor.sh with your node hostnames.

Author

nguyenpanda
Email: [email protected]
Website: https://nguyenpanda.com

License

MIT License