net_mdev: Fast path user space I/O
summary
net_mdev ports the kernel mediated-device idea to NICs so user space handles the fast path while the kernel keeps control ops. Prototype uses the existing VFIO mediated device model (Linux 4.10+), needs ~500 LOC kernel changes and ~1,000 LOC user driver instead of 30–40k, supports slot/slope/tape buffer models and zero copy. Measured results: 10 GbE ≈ 13 Mpps TX / 8 Mpps RX, 40 GbE ≈ 48 Mpps TX and 10.3 Mpps RX (single core), approaching DPDK performance in tests.
AI-written summary · about
chapters (17)
- 0:00Introduction and motivation
- 3:19Kernel limitations and VFIO approach
- 7:15Mediated device framework benefits
- 10:43Performance goals and hardware support
- 14:13Raw performance results
- 17:44Kernel vs user space I/O design
- 23:10Transition to user space operation
- 27:42User space implementation steps
- 31:57Future improvements and challenges
- 38:14Q&A: Virtual machines and security
- 41:02Q&A: Flow control and interrupts
- 44:32Q&A: Performance and CPU usage
- 47:20Q&A: Interface configuration
- 50:14Q&A: Security and vendor collaboration
- 53:31Q&A: Hardware design and trade-offs
- 57:57Q&A: Community adoption and upstreaming
- 1:00:25Q&A: PCIe transaction overhead
related talks
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
43mLessons Learnt from Reusing QDMA NIC to Base Band PMD
51mAccelerate containers and VMs networking with vDPA
31mHow to capture 100G Ethernet traffic at wire speed to local disk
52mIn the Kernel Trenches: Mastering Ethernet Drivers on Linux
24mMetron: NFV Service Chains at the True Speed of the Underlying Hardware
15mCharacterizing Off-path SmartNIC for Accelerating Distributed Systems