Controlling Adaptive HARQ Erasure Coding for Real-Time Transport Under Channel Model Mismatch

Abstract

Networking is an essential component of real-time cyber-physical systems, and hence, networking problems are increasingly real-time problems. As these systems expand beyond single, controlled links, they require predictably reliable end-to-end communication primitives, effectively transferring their deadline constraints to the underlying multi-hop network. Standard transport protocols often optimize either only for latency (e.g., UDP) or for full reliability (e.g., TCP, QUIC), with the latter potentially leading to unbounded retransmission delays. Partial reliability bridges the gap between these extremes, offering configurable reliability within a bounded delay. To remain bandwidth-efficient and robust to dynamic network conditions, these schemes must abandon static forward error correction (FEC) and move toward adaptive loss recovery. A fundamental challenge is the model mismatch problem: real-time adaptive loss recovery schemes require simple, efficiently interpretable models to estimate network conditions, yet these models systematically underfit complex real-world network dynamics.
To provide robustness against this inherent underfitting, we present a closed-loop control architecture that applies an adaptive safety margin to the network estimates, ensuring the system meets a deadline-constrained reliability target. We frame this contribution within the Predictably Reliable Real-time Transport protocol (PRRT), which allows applications to configure a loss and deadline constraint. The control system observes the packet delivery deficit (packet debt) to quantify the extent to which the end-to-end packet-loss rate deviates from the application’s target loss rate. A compensated loss rate is then fed into a novel constraint-aware, anytime incremental search algorithm that derives a near-optimal Hybrid Automatic Repeat Request (HARQ) coding configuration that combines the benefits of proactive and reactive packet-loss recovery to satisfy the application’s loss and delay constraints. New to this search is its awareness of the encoding and decoding complexity of the resulting coding configurations. This allows devices to adaptively limit the search space to configurations within their computational capabilities, which is essential for constrained edge devices. We provide a new high-performance Rust reference implementation of PRRT and demonstrate that the system converges towards the target loss rate, even under model mismatch, while also quickly adapting to shifts in network conditions.