Architecture & Performance¶
Omnipkg is designed to solve the “Linker Lock” problem—the inability to load multiple versions of C-extension libraries (like TensorFlow or NumPy) in a single Python process.
It achieves this through a distributed Worker Daemon architecture that outperforms traditional virtualization methods (Docker, Conda) for Python-specific workflows.
Core Documentation¶
-
- Understand the Manager-Worker architecture.
- Learn how Universal CUDA IPC achieves zero-copy data transfer.
- See the decision matrix for Legacy Loader vs. Daemon.
-
- See how omnipkg achieves 2ms execution latency.
- Comparison against Docker, Conda, and Venv.
- Memory efficiency analysis (1.9x vs Docker).
Key Concepts¶
1. Process Isolation > Virtualization¶
Instead of virtualizing the entire OS (Docker) or just the filesystem (Conda), omnipkg virtualizes the Python Runtime. This allows it to strip away overhead while maintaining the strict ABI isolation required for AI frameworks.
2. Intelligent Dispatch¶
The daemon automatically routes payloads based on data locality: * GPU Data? $\to$ Universal CUDA IPC (<5µs) * Large CPU Data? $\to$ Shared Memory Ring Buffer (~5ms) * Config Data? $\to$ Standard JSON Sockets (~10ms)