LossLens: Robust Byte-level TCP Loss Detection at the Datacenter Gateway

Datacenter gateway is the key vantage point for measuring north–south traffic. Among gateway measurement tasks, TCP loss detection is the most crucial one, since it reveals the quality of service (QoS) and helps localize losses to upstream or downstream network fabrics. However, previous loss detection systems were primarily designed for datacenters and the WAN. They fall short in this scenario due to three gateway-specific challenges: single-vantage constraint, lack of explicit loss signals, and highly noisy traffic. To this end, we present LossLens, a byte-level loss detection system for the high-availability physical gateway. For non-intrusiveness, it adopts a tap device to mirror and process packets off-path on a P4-capable switch. For robustness, it efficiently tracks top-k TCP sequence ranges of heavy hitters in the programmable data plane. It exploits the data and control planes cooperatively to denoise traffic and identify actual losses. Testbed experiments and trace-driven simulations show that LossLens achieves accurate, robust, and efficient loss detection and outperforms prior approaches.