In edge computing, efficient GPU utilization is critical—but legacy protocols like TCP are hitting their limits. In our recent study presented at WiMob 2024, we investigate GPU virtualization in edge scenarios using the QUIC protocol as a next-gen alternative to TCP.
🔍 Why this matters:
Traditional TCP-based GPU sharing over the network suffers from latency, head-of-line blocking, and lack of stream control—especially problematic in wireless or constrained environments.
🔧 Our experimental setup emulated real-world edge conditions (1 Gbit/s link, 0.01% loss, 1ms delay) and tested GPU API remoting using both TCP and QUIC.
📈 Key results:
- QUIC reduced Peak Age of Information (pAoI) from 2s (TCP) to just 400ms, enabling near real-time updates.
- Concurrent streams, prioritization, and asynchronous API calls in QUIC significantly improved performance.
- Stream cancellation, on the other hand, degraded freshness—highlighting nuanced protocol behavior.
The findings suggest QUIC is a viable enabler for low-latency, remote GPU access at the edge—critical for machine learning, analytics, and real-time AI applications.
📄 Read the full paper: https://zenodo.org/records/14717463
🌐 Project info: cleverproject.eu