open-mps

Open-source reimplementation of NVIDIA MPS (Multi-Process Service) for GPUs where the official MPS runtime is unavailable or unsupported.

Overview

open-mps aims to provide a lightweight open-source alternative to NVIDIA MPS, enabling multiple processes to share a single GPU efficiently.

The project is primarily targeted at:

• Legacy GPUs where NVIDIA MPS is not available
• Unsupported environments
• Experimental GPU scheduling research
• Open-source GPU runtime experimentation

By reimplementing key MPS concepts, open-mps allows concurrent GPU workloads to be multiplexed through a shared execution layer.

Motivation

NVIDIA’s Multi-Process Service improves GPU utilization by allowing multiple CUDA processes to share a GPU context and schedule work concurrently.

However, official MPS support is limited by:

• GPU architecture requirements
• driver restrictions
• proprietary implementation
• lack of support for older GPUs

open-mps attempts to recreate similar functionality in an open and portable way, enabling experimentation and broader hardware compatibility.

Goals

• Provide multi-process GPU execution sharing
• Support legacy or unsupported NVIDIA GPUs
• Offer a fully open-source implementation
• Enable research on GPU scheduling and virtualization
• Maintain minimal runtime overhead

Planned Architecture

High-level design:

+-------------------+
| Client Process A  |
+-------------------+
          |
+-------------------+
| Client Process B  |
+-------------------+
          |
          v
+---------------------------+
| open-mps daemon / broker  |
|                           |
| - request queue           |
| - context manager         |
| - scheduler               |
+---------------------------+
          |
          v
+-------------------+
| CUDA Driver / GPU |
+-------------------+

The open-mps daemon will:

• manage GPU contexts
• multiplex kernels from multiple processes
• schedule execution
• coordinate memory and streams

Current Status

🚧 Early development

This repository currently focuses on:

• architecture design
• core runtime prototype
• GPU request multiplexing

Expect breaking changes while the system evolves.

Roadmap

Planned features:

• basic daemon runtime
• process registration
• kernel dispatch proxy
• CUDA stream multiplexing
• memory sharing layer
• fair GPU scheduling
• debugging and metrics

Long-term ideas:

• GPU sandboxing
• container integration
• scheduler plugins
• experimental GPU virtualization

Non-Goals (for now)

• full CUDA runtime replacement
• binary compatibility with official MPS
• support for all CUDA features

Contributing

Contributions are welcome.

Areas where help is especially useful:

• CUDA runtime internals
• GPU scheduling algorithms
• driver-level experimentation
• testing across GPU generations

License

MIT License

Disclaimer

open-mps is an independent open-source project and is not affiliated with or endorsed by NVIDIA.