Explore a groundbreaking congestion control algorithm for datacenter networks in this 16-minute conference talk from NSDI '22. Dive into PowerTCP, a novel approach that leverages in-band network telemetry to achieve fine-grained congestion control by adapting to the bandwidth-window product. Learn how this algorithm outperforms existing solutions in both traditional and emerging reconfigurable datacenter topologies, significantly reducing tail flow completion times and improving circuit utilization. Discover the key concepts behind PowerTCP, including its fast reaction time, ability to handle dynamic network environments, and bursty traffic patterns. Examine the analytical and empirical evidence supporting PowerTCP's superior performance, and gain insights into its potential to push the performance limits of datacenter networks.
PowerTCP - Pushing the Performance Limits of Datacenter Networks
Overview
Syllabus
Intro
Brief context of electrical systems
Analogy to networked systems
PowerTCP in a Nutshell
Fear of the buffer Buffer per unit capacity (KB/Gbps)
Timeline of congestion control in datacenters
Problems of existing approaches
Summary of Our Analysis Voltage-based
The notion of power
PowerTCP control law
Evaluation - Incast
Evaluation - Fairness & Stability
Evaluation - Workload PowerTCP
Evaluation - Reconfigurable Networks
Conclusion
Taught by
USENIX
Related Courses
-
Adapting TCP for Reconfigurable Datacenter Networks
-
Backpressure Flow Control
-
Datacenter RPCs Can Be General and Fast
-
Poseidon - Efficient, Robust, and Practical Datacenter CC via Deployable INT
-
Orca - Server-assisted Multicast for Datacenter Networks
-
Shoal - A Network Architecture for Disaggregated Racks
