Collateral damage: the impact of optimised TCP variants on real-time traffic latency in consumer broadband environments

In recent years a number of TCP variants have emerged to optimise some aspect of data transport where high delay-bandwidth product paths are common. We evaluate a different scenario - latency-sensitive UDP-based traffic sharing a consumer-grade 'broadband' link with one or more TCP flows. In particular we compare Linux implementations of NewReno, H-TCP and CUBIC. We find that dynamic latency fluctuations induced by each TCP variant is a more significant differentiator than 'goodput' (useful throughput), and that CUBIC induces far more latency than either H-TCP or NewReno when multiple TCP flows are active concurrently. This potential for 'collateral damage' should influence future efforts to re-design TCP for widespread deployment