NetKernel: Making Network Stack Part of the Virtualized Infrastructure
summary
NetKernel decouples the TCP/IP stack into an operator-run network stack module implemented as a VM, redirecting BSD socket calls via a guest shim into lightweight 32-byte NQE metadata queues plus huge pages for payload and a lockless queue engine. Testbed with a Mellanox 100G NIC shows line-rate 100 Gbps throughput, shared-memory inter-VM paths >2× for 8 KB messages, and mTCP deployed unchanged yielding ~1.8× RPS on 4 vCPU workloads. Some mTCP runs fail on certain core counts and multi-tenant kernel filters remain hard.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:30Current architecture in the cloud
- 0:59Motivation: Tenants
- 1:55Motivation: Provider
- 3:47Making Network Stack Part of the Virtualized Infrastructure
- 4:05Benefits
- 4:54Design Challenges
- 5:35Transparent socket API redirection
- 6:33A lightweight semantics channel
- 7:46Scalable lockless queues
- 8:23VM based NSM.
- 9:18Netkernel
- 10:48Implementation
- 11:16Use Cases #1: Multiplexing
- 13:51Use Case #3: Shared Memory Networking
- 14:37Microbenchmarks: Throughput
- 15:14Microbenchmarks: RPS
- 15:56Discussion and future directions
- 19:14Recap
related talks
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
14mHigh-throughput and Flexible Host Networking for Accelerated Computing
17mRearchitecting the TCP Stack for I/O-Offloaded Content Delivery
01h 01mnet_mdev: Fast path user space I/O
27mHigh speed load balancing from the Linux Kernel
24mNetBlocks: Staging Layouts for High-Performance Custom Host Network Stacks
18mNICA: An Infrastructure for Inline Acceleration of Network Applications
31mHow to capture 100G Ethernet traffic at wire speed to local disk