L, Q, R, and T -- Which Spin Bit Cousin Is Here to Stay? (2106.13710v1)

Abstract: Network operators utilize traffic monitoring to locate and fix faults or performance bottlenecks. This often relies on intrinsic protocol semantics, e.g., sequence numbers, that many protocols share implicitly through their packet headers. The arrival of (almost) fully encrypted transport protocols, such as QUIC, significantly complicates this monitoring as header data is no longer visible to passive observers. Recognizing this challenge, QUIC offers explicit measurement semantics by exposing the spin bit to measure a flow's RTT. Ongoing efforts in the IETF IPPM working group argue to expose further information and enable the passive quantification of packet loss. This work implements and evaluates four currently proposed measurement techniques (L-, Q-, R-, and T-bit). We find that all techniques generally provide accurate loss estimations, but that longer algorithmic intervals for Q and R, yet foremost for T, complicate detecting very small loss rates or loss on short connections. Deployment combinations of Q & R as well as Q & L, thus, have the best potential for accurately grasping the loss in networks.