Evaluating Stream Mapping Strategies for MQTT over QUIC

Read full text Summarise with VIRAT
Authors: Fabrício Bracht
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

MQTT, the dominant IoT messaging protocol, is conventionally carried over TCP, where head-of-line (HOL) blocking couples logically independent topic flows: a single packet loss delays delivery across all topics sharing a connection. Three stream mapping strategies, namely control-only, per-topic, and per-publish, each offer a different point in the isolation–overhead tradeoff space for MQTT over QUIC. We implement them in mqtt5, an open-source MQTT library whose broker-autonomous per-topic delivery-stream mapping is absent from existing brokers, and evaluate them across four experiments on GCP infrastructure under controlled network impairment. Using a density-corrected spike-isolation metric and per-topic tail latency to quantify HOL blocking, we find that the transport, not stream mapping, sets the level of HOL blocking at heavy (5%) loss: all three strategies keep the median per-topic tail below TCP at 5% loss (control-only 132 ms versus TCP's coupled 351 ms). Multistream adds spike-timing isolation, but as a bounded tradeoff rather than a free benefit: per-topic's isolation holds only at small topic counts, reaching single-stream coupling from about four topics, while per-publish stays isolated, at a tail latency that depends on loss and rate. On a single connection the strategies differ ninefold in publish rate, but we show that gap is set by two QUIC transport constants, the concurrent stream limit and the per-stream receive window, rather than by the mappings themselves. At our 8-topic, 500 msg/s reference point, control-only, adding the least packet overhead, has the lowest median and 90th-percentile tail among the QUIC strategies at 5% loss; at 2% it is indistinguishable from per-topic and at 1% it is the highest of the three, and at 125 msg/s under 5% loss per-publish is lowest, so which strategy has the lowest tail depends on both loss and rate. Under loss, QUIC delivers 1.26× the throughput of an encryption-matched TCP+TLS 1.3 baseline at 1% loss, rising to 1.78× at 10%, once the emulated loss removes single packets; applying the same loss per kernel buffer on the sending host inflates that advantage to 2–3×. These results provide concrete strategy selection guidelines for deploying MQTT over QUIC in IoT environments.

Fields of Study & Subjects

Network Traffic and Congestion Control IoT Networks and Protocols Mobile and Web Applications

Publication details

Journal / source
SSRN Electronic Journal
Publication date
1 January 2026
Language
EN
Document type
Preprint