Revisiting delay-based TCP

Side-bar: “newtcp” project at CAIA

Loss-based TCP – cyclical queue filling

Induced queuing delay – “Collateral damage”

Why delay-based TCP?

Our first foray – “delay threshold” TCP

Delay-gradient TCP: Less configuration

Delay gradient as a congestion signal

Inferring queue states

Inferring queue states

Window Progression when no packet loss

Probabilistic 'back-off'

An aside: our testbed

CDG induces far less queuing delay

Reaction to packet loss

CDG's improved loss tolerance

Competing with loss-based flows #1

Competing with loss based flows #2

CDG achieves a reasonable goodput

CDG achieves better goodput with loss

Sharing between CDG and NewReno

Sharing between flows with different RTT