Replicating and Extending SUSS: TCP Slow-Start Acceleration on AWS

Abstract

TCP slow-start's conservative ramp-up leaves short flows well below available bandwidth during their most performance-critical rounds. SUSS (SIGCOMM 2024) addresses this with a sender-side add-on to CUBIC that predicts when exponential growth can safely continue and accelerates the congestion window accordingly, reporting over 20% flow-completion-time (FCT) improvement across 28 diverse scenarios. We replicate SUSS on an AWS-based testbed spanning six wired inter-region paths with RTTs from 8 to 262 ms and one residential WiFi path. We further extend the original paper in two directions. First, we characterize the relationship between SUSS's benefit and RTT when the server region and last-hop link type are held fixed. We find a clear lower-bound effect below 50 ms RTT — the regime where the original paper's claim does not apply — and substantial path-specific variation above it. Across our five wired scenarios where the claim applies, improvement averaged across flow sizes ranges from 27% to 44%, with no monotonic relationship to RTT. RTT alone is not a reliable predictor of SUSS's benefit. Second, we adapt SUSS's acceleration principle to BBR's STARTUP phase as an exploratory extension, finding that the principle transfers to a fundamentally different congestion control algorithm but at substantially smaller improvement magnitude (1–9% versus 12–47%). We believe this is consistent with BBR's already-aggressive baseline that leaves a narrower under-utilization window for acceleration to exploit. We additionally evaluate SUSS under concurrent long-lived background traffic and find that its FCT improvements persist on contended paths under both CUBIC and BBR.