A large-scale deployment of DCTCP
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
DCTCP rollout across thousands of racks reduces retransmissions by about 75% drop in retransmissions and cuts one latency-sensitive service’s P90/P99 by 38% P90/P99 reduction, but deployment surfaces practical problems: ECN-marking missed for SYNs causing SYN retransmits, switch queue layout forcing mixed CCAs on the same port, and incast-induced stalls. Solution stack uses an eBPF CCA selector plus a user-space iterator to flip CCAs, a receiver flow-control knob for incast, and a kernel fix to read ECT from the requesting socket.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 0:15DCTCP basics and background
- 2:42Performance results
- 3:28Rollout challenges and design
- 5:19Solving deployment issues
related talks
19mTackling DC Congestion and Bursts
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
22mAdapting TCP for Reconfigurable Datacenter Networks
31mTuning Linux TCP for data-center networks
22mLIVE Q&A #3
23mACK to the Future: TCP in 2025
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
16mR-TCP: A Framework to Optimize TCP Performance Over Rate-Limiting Networks