Remote TCP Connection Offload and Applications
summary
XO enables direct server return by serializing TLS/TCP connection state and restoring it on backends, cutting L7 proxy CPU and bandwidth. It installs per-connection redirect rules with an EBPF+TC hybrid redirection (works on ConnectX‑5 and Agilio) and uses a cancelable user-space TC queue to avoid stale hardware rules. Evaluation shows 27–365% higher throughput vs conventional L7LBs and up to 135% and 300% gains for Ceph and nginx respectively.
AI-written summary · about
chapters (16)
- 0:00Remote TCP connection offload intro
- 2:26XO reduces CPU/network overhead
- 3:54Limitations of L7 load balancer optimizations
- 4:37XO architecture and connection migration
- 6:26eBPF and TC packet redirection
- 8:12Cancelable queue for hardware rules
- 10:02Porting apps to XO
- 11:51Kernel network stack advantages
- 12:11Performance vs conventional proxy
- 14:49CPU impact on LB and servers
- 15:30State transfer latency
- 16:40TCP state serialization optimization
- 17:27Throughput and CPU in EnginX/SAP
- 18:10Deploying XO in virtualized envs
- 19:16Future work on congestion control
- 19:39Summary and follow-up
related talks
21mAccelTCP: Accelerating Network Applications with Stateful TCP Offloading
14mFlexTOE: Flexible TCP Offload with Fine-Grained Parallelism
17mRearchitecting the TCP Stack for I/O-Offloaded Content Delivery
19mConntrack offload: why and how
42mAccelerating Envoy with the Linux Kernel
15mHA/TCP: A Reliable and Scalable Framework for TCP Network Functions
21mTCP ≈ RDMA: CPU efficient Remote Storage Access with i10
17mOpening Up Kernel-Bypass TCP Stacks