Scaling Flexicast QUIC for Large-Scale Software Distribution
Authors: Louis Navarre, Olivier Bonaventure
Published: Preprint, under submission, 2026
Abstract:
Large-file distributions, such as video game patches and OS updates, generate traffic spikes, forcing ISPs to over-provision their networks beyond daily requirements. Content Delivery Networks usually deliver these files using unicast, generating large amounts of duplicate traffic. We reconsider multicast by building upon Flexicast QUIC (FCQUIC), a QUIC extension that combines efficient multicast delivery with robust per-receiver unicast paths, offering fallback when needed. We address two challenges related to large-scale multicast file transfer. First, the ACK implosion problem: as the number of receivers grows, acknowledgment traffic can overwhelm the source. We derive an adaptive acknowledgment strategy in which the source dynamically advertises an acknowledgment delay, capping the aggregate rate from the receivers below a defined threshold. Second, loose receiver synchronization: because receivers request the same file in concentrated but not simultaneous bursts, we design a file-rolling system which splits the file into blocks and sends them in a carousel, allowing late receivers to begin processing the application at the next block rather than waiting for a complete retransmission. We implement both mechanisms in FCQUIC and evaluate them on CloudLab with up to 1000 receivers. Our adaptive strategy caps the acknowledgment rate without degrading per-receiver goodput and outperforms both QUIC and NORM, a multicast state-of-the-art protocol, with numerous receivers across numerous scenarios, including heterogeneous receiver arrival and packet loss. We further model the CPU load of the FCQUIC source and show that the dominant factor is the acknowledgment delay, which we can control, allowing us to consider larger-scale deployments.
Recommended citation: Navarre, L., & Bonaventure, O. (2026). Scaling Flexicast QUIC for Large-Scale Software Distribution..
Download Paper | Download Bibtex
