Halfback: running short flows quickly and safely

Abstract

Interactive applications like web browsing are sensitive to latency. Unfortunately, TCP consumes significant time in its start-up phase and loss recovery. Existing sender-side optimizations use more aggressive start-up strategies to reduce latency, but at the same time they harm safety in the sense that they can damage co-existing flows' performance and potentially the network's overall ability to deliver data. In this paper, we experimentally compare existing solutions' latency performance and more importantly, the trade-off between latency and safety at both the flow level and the application level. We argue that existing solutions are still operating away from the sweet spot on this trade-off plane. Based on the diagnosis of existing solutions, we introduce Halfback, a new short-flow transmission mechanism that operates on a better latency-safety trade-off point: Halfback achieves lower latency than the lowest latency previous solution and at the same time significantly better safety. As Halfback is TCP-friendly and requires only sender-side changes, it is feasible to deploy.