Opening Up Kernel-Bypass TCP Stacks
summary
Six kernel and kernel-bypass TCP stacks are compared on identical hardware and the same HTTP workload with per-stack optimized servers, disabled NIC offloads, and implemented-missing TCP options to make a fair test. No single stack wins every metric: Linux wins large-transfer throughput, Demi achieves lowest idle RPC latency, Ex gives highest RPC throughput, TAS scales best across many cores. mTCP and Ex show good connection scalability. Reproducing and profiling bypass stacks proves surprisingly hard, and several stacks lack generality for mixed workloads.
AI-written summary · about
chapters (18)
- 0:00Kernel-bypass TCP stacks intro
- 1:01Why TCP dominates despite flaws
- 2:08Criticisms of kernel TCP
- 3:14User-space TCP stacks emerge
- 4:42Challenges in practical TCP stacks
- 5:45Lack of bypass stack studies
- 6:48Paper’s contribution and method
- 8:33Tested bypass stacks overview
- 10:19Performance setup and fairness
- 11:01Throughput results for large data
- 11:23Latency and RPC performance
- 12:06Connection scalability tests
- 12:28Multi-core scalability analysis
- 13:11Profiling bypass stacks challenges
- 13:57Future network stack design
- 15:25Run-to-completion model limits
- 16:07Key takeaways and conclusion
- 16:50Q&A
related talks
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
17mRearchitecting the TCP Stack for I/O-Offloaded Content Delivery
28mMultipath TCP hands-on on Fedora
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure
25mEmpowering High Performance User Space Network Stack on Arm
21mAccelTCP: Accelerating Network Applications with Stateful TCP Offloading
21mRemote TCP Connection Offload and Applications
13mRakaia: Scalable In-Kernel Scheduling for TCP-Based RPCs