Tackling DC Congestion and Bursts
summary
Deploys DCTCP with ECN for intra-region flows using a custom cgroup BPF selector, fixes a kernel SYN-ECN bug with a temporary BPF marker, and measures adoption (short flows switch in milliseconds, long-lived take weeks). Observes a 91.5% drop in rack switch discards and large retransmission reductions tied to DCTCP. Builds egress BPF receiver window tuning with a host-side threshold estimator, cutting retransmissions by ~80–90%, queue utilization by 50%, and SRRT by 30%. Future work: unify signals and fleet-safe window caps.
AI-written summary · about
chapters (14)
- 0:00DC congestion and bursts
- 1:49Data Center TCP signals
- 3:15Intra vs inter-region traffic
- 4:40DCTCP deployment methods
- 5:46Kernel bug and SYN fix
- 6:33DCTCP adoption and impact
- 8:42Microburst challenges
- 10:32Microbursts and TCP windows
- 12:37In-cast problem and bottlenecks
- 14:01Receiver window tuning
- 15:50Optimal window thresholds
- 16:54Control plane for tuning
- 17:14Feed infra service results
- 17:55Next steps and improvements
related talks
14mA large-scale deployment of DCTCP
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
16mPowerTCP: Pushing the Performance Limits of Datacenter Networks
23mACK to the Future: TCP in 2025
31mTuning Linux TCP for data-center networks
22mAdapting TCP for Reconfigurable Datacenter Networks
28mSafe sharing of the network with eBPF
16mR-TCP: A Framework to Optimize TCP Performance Over Rate-Limiting Networks