PowerTCP: Pushing the Performance Limits of Datacenter Networks
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Introduces PowerTCP, a congestion control that drives send-window updates from the bandwidth–window product (called power) and measures it via in-band network telemetry (with an RTT-only approximation). Implements multiplicative increase/decrease plus EWMA and evaluates on RDMA oversubscribed fabrics and reconfigurable datacenters. Claims one-RTT convergence, cuts short-flow tail FCTs by ~80% vs DCQCN/TIMELY and ~33% vs HPCC, reaches ~85% circuit utilization in reconfigurable setups while keeping queues short.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:32Brief context of electrical systems
- 0:51Analogy to networked systems
- 1:20PowerTCP in a Nutshell
- 3:00Fear of the buffer Buffer per unit capacity (KB/Gbps)
- 4:25Timeline of congestion control in datacenters
- 5:43Problems of existing approaches
- 7:11Summary of Our Analysis Voltage-based
- 8:04The notion of power
- 9:32PowerTCP control law
- 10:49Evaluation - Incast
- 12:02Evaluation - Fairness & Stability
- 13:10Evaluation - Workload PowerTCP
- 14:17Evaluation - Reconfigurable Networks
- 15:41Conclusion
related talks
19mTackling DC Congestion and Bursts
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
31mTuning Linux TCP for data-center networks
23mACK to the Future: TCP in 2025
22mAdapting TCP for Reconfigurable Datacenter Networks
19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable Switches
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
14mA large-scale deployment of DCTCP
