Rearchitecting the TCP Stack for I/O-Offloaded Content Delivery
summary
Claims a split TCP stack called IO-TCP that keeps control on the host CPU and pushes high-throughput data work to SmartNICs, using p2p PCIe DMA so the NIC reads NVMe directly and sends packets. Prototype modifies the host TCP and runs a SmartNIC TPDK app with four offload APIs offload_open offload_read offload_fstat offload_close, ports lighttpd with ~10 lines changed, saturates four NVMe drives at ~80 Gbps with one CPU core, and shows 1.8x–2.1x throughput gains versus Atlas and Linux TCP.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:49Computing Hardware Development Trend
- 1:50CPU Consumption for HTTP Video Streaming Server
- 4:14Challenges in Operating TCP Stack
- 5:0210-TCP: Split TCP Stack Architecture for Content Delivery
- 6:3310-TCP based Web Server Workflow
- 7:44Key Operation for Offloading: SEND Command
- 9:14How to Handle Retransmission?
- 10:03Implementation
- 10:46Experiment Setup
- 13:37Source of Performance Improvement
- 15:40Summary
related talks
21mAccelTCP: Accelerating Network Applications with Stateful TCP Offloading
21mTCP ≈ RDMA: CPU efficient Remote Storage Access with i10
14mHigh-throughput and Flexible Host Networking for Accelerated Computing
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure
14mFlexTOE: Flexible TCP Offload with Fine-Grained Parallelism
26mApplying New High-Throughput Network Solutions in IEC (Integrated Edge Cloud)
39mSecured I/O in the Cloud with NVMe-Over-TCP