Public thesis defense Louis NAVARRE - ICTEAM
sst |
Improving the Efficiency and Flexibility of Network Communications
Wednesday August 26, 2026, 4pm - Auditorium BARB94 - Place Sainte-Barbe 1, 1348 Louvain-la-Neuve
Modern networks are provisioned for peak load rather than average load, and the ratio between the two continues to grow. This over-provisioning is largely driven by sporadic one-to-many events -- live streaming, software updates, or popular game releases -- during which many receivers request the same content at nearly the same time. This thesis aims to improve the efficiency of these networks, primarily by reconsidering the use of multicast to reduce the gap between average and peak traffic.
A first consequence of this over-provisioning is the spare bandwidth that remains unused most of the time. We leverage it to reduce application latency: instead of recovering from packet losses through retransmissions, which add at least one round trip, we transparently protect traffic with Forward Erasure Correction within a tunnel deployed at the network layer, without requiring end hosts to implement the mechanism themselves. We show that a carefully designed software implementation operates at high speed and reduces application tail latency over a real network that exhibits losses unrelated to congestion.
The core of this thesis reconsiders multicast, the natural way to deliver identical data to many receivers, thereby directly reducing the peak traffic generated by these one-to-many events. Despite its efficiency, multicast is not available everywhere on the Internet, and the transport protocols historically designed for it assume a source that scales independently of the number of receivers. Building on QUIC and its multiple paths management extension, we design Flexicast QUIC, which combines shared multicast delivery with per-receiver unicast paths within a single connection. Receivers seamlessly fall back to unicast whenever multicast fails or underperforms, turning multicast from an all-or-nothing decision into an optimization: multicast when available, unicast otherwise. Because the source now knows the exact set of active receivers, we further revisit the historical acknowledgment-implosion problem in reliable multicast, pacing receiver feedback to keep Flexicast QUIC scalable as the group grows.
Finally, this work led us to analyze the optimistic acknowledgment attack in QUIC, in which a malicious receiver acknowledges data it has not yet received, inflating the sender's transmission rate. We show that numerous implementations were exposed to it, even though QUIC's specification already describes a protection against it as an optional feature. Our analysis led several widely deployed implementations to finally adopt this protection.
Jury members
Prof. Olivier Bonaventure (UCLouvain), Supervisor
Prof. Charles Pecheur (UCLouvain), Chairperson
Prof. Ramin Sadre (UCLouvain), Secretary
Prof. Etienne Riviere (UCLouvain)
Prof. Jörg Ott (TUM, Germany)
Prof. Matthias Wählisch (TU Dresden, Germany)