G-NET: Effective GPU Sharing in NFV Systems
summary
G-NET implements a hypervisor manager that creates a common GPU context to enable spatial GPU sharing across NFV containers, uses a service-chain scheduler that assigns streaming multiprocessors and batch sizes, and enforces GPU memory isolation with an ISO pointer plus a zero-copy shared buffer. Built with Docker and DPDK, it reports up to 70.8% throughput improvement and up to 44.3% latency reduction versus temporal GPU virtualization, showing the design both practical and effective.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:28Network Function Virtualization (NFV)
- 1:36GPUs in Accelerating Network Functions
- 3:04GPU-Accelerated Network Functions
- 3:55Why GPUs Have not Been Utilized in NFV Systems?
- 4:40Current Way of GPU Virtualization is inefficient
- 6:02Hyper-Q for Spatial GPU Sharing
- 8:33G-NET: System Workflow
- 9:56Achieve Predictable Performance
- 11:43Service Chain Based Scheduling
- 13:52IsoPointer for GPU Memory Isolation
- 14:28NF Development
- 15:38Evaluation
- 16:56Latency
- 17:25Conclusion
related talks
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
19mFpgaNIC: An FPGA-based Versatile 100Gb SmartNIC for GPUs
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure
51mAccelerate containers and VMs networking with vDPA
17mEnabling Efficient GPU Communication over Multiple NICs with FuseLink
14mHigh-throughput and Flexible Host Networking for Accelerated Computing
13mLive Q&A Session #2 – Shashi Gandham
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...