ACK to the Future: TCP in 2025
summary
Covers concrete tuning: shows Cubic collapsing under 0.05% random loss while BBR tolerates loss and probes RTTs, notes BBR v3 out of tree versus mainline BBRv1, and recommends enabling ECN if network devices set CE bits. Demonstrates measuring windows with Wireshark and explains the magic bandwidth‑delay product — e.g. a 1 Gbit Dublin–Seattle link needs ~17.12 MB of in‑flight data — and reviews NIC offloads like segmentation offload and checksum offload, warning about Intel 700 TCPOQ reordering spikes (330M) that break performance.
AI-written summary · about
chapters (16)
- 0:00Welcome
- 0:39TCP’s 50-year legacy
- 3:12Congestion control fundamentals
- 6:26BBR vs Cubic performance
- 7:31Why Google moved to Quick
- 8:56BBR v3 and kernel lag
- 9:39Explicit Congestion Notification
- 11:05TCP receive window sizing
- 14:17Bandwidth delay product
- 16:46Buffer sizing mistakes
- 17:06Hardware offload for TCP
- 18:10Checksum and segmentation offload
- 19:56Flow affinity and queues
- 20:39Monitoring TCP health
- 21:00Hardware offload pitfalls
- 22:25Summary
related talks
19mTackling DC Congestion and Bursts
22mAdapting TCP for Reconfigurable Datacenter Networks
16mR-TCP: A Framework to Optimize TCP Performance Over Rate-Limiting Networks
31mTuning Linux TCP for data-center networks
22mLIVE Q&A #3
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
45mReducing Network Latency: Innovations for a Faster Internet