Scheduler-Agnostic Adaptive-FEC for MPQUIC: Field Evaluation over Commercial Cellular Paths

Reusing QUIC loss signals and Reed-Solomon coding to harden MPQUIC Datagram against vehicular packet loss without touching the scheduler

Takuma Tsubaki; Soto Anno; Seiya Komatsu; Takashi Torii; Takuya Tojo· 2026· DOI 10.48550/arXiv.2607.14482

The core problem

Multipath transport has long been studied as a means of improving robustness over heterogeneous networks, and Multipath QUIC (MPQUIC) is particularly attractive for low-latency applications because QUIC Datagram enables transmission without retransmissions. That advantage, however, carries a key drawback: QUIC Datagram is vulnerable to packet loss, especially in vehicular cellular environments characterized by handovers and rapidly varying radio conditions. Prior work has explored adaptive-FEC and coded multipath designs, but practical deployment remains challenging when such mechanisms demand substantial modifications to the transport stack or the scheduler. This paper addresses that deployment gap by presenting an adaptive Forward Error Correction (FEC) scheme for MPQUIC Datagram that reuses QUIC loss detection signals together with an off-the-shelf Reed-Solomon (RS) coding library. The central research question is whether scheduler-agnostic adaptive FEC can deliver practical latency and reliability gains under real commercial cellular conditions without invasive changes to the MPQUIC stack.

Under the main 4.0 Mbps setting, the proposed scheme improved both reliability and latency compared with No-FEC. The average one-way delay was reduced from 103.0 ms to 70.8 ms, and the 95th-percentile delay fell from 281.2 ms to 142.3 ms. The packet loss rate dropped from 1.7% to 0.8%, achieved with an average coding rate of 0.94. These figures were obtained across three commercial LTE/5G paths in vehicular field experiments, meaning the gains reflect real handover behavior and rapidly varying radio conditions rather than simulation. The tail-latency improvement is especially notable: the 95th-percentile delay was roughly halved, which matters for low-latency applications that depend on QUIC Datagram and cannot rely on retransmissions.

Multipath transport has long been studied as a means of improving robustness over heterogeneous networks, and Multipath QUIC (MPQUIC) is particularly attractive for low-latency applications because QUIC Datagram enables transmission without retransmissions. That advantage, however, carries a key drawback: QUIC Datagram is vulnerable to packet loss, especially in vehicular cellular environments characterized by handovers and rapidly varying radio conditions. Prior work has explored adaptive-FEC and coded multipath designs, but practical deployment remains challenging when such mechanisms demand substantial modifications to the transport stack or the scheduler. This paper addresses that deployment gap by presenting an adaptive Forward Error Correction (FEC) scheme for MPQUIC Datagram that reuses QUIC loss detection signals together with an off-the-shelf Reed-Solomon (RS) coding library. The central research question is whether scheduler-agnostic adaptive FEC can deliver practical latency and reliability gains under real commercial cellular conditions without invasive changes to the MPQUIC stack.

The proposed scheme operates entirely on signals already exposed by QUIC, avoiding any scheduler-specific hooks. The sender estimates smoothed per-path loss rates from QUIC loss signals and adaptively determines the number of parity packets for each FEC block according to the expected packet loss. Parity generation and recovery rely on a standard Reed-Solomon coding library, keeping the implementation lightweight and portable.

The results indicate that scheduler-agnostic adaptive-FEC can provide practical latency and reliability gains for MPQUIC Datagram under measured vehicular mobile conditions. Two design properties explain the outcome. First, by reusing QUIC loss detection signals, the scheme avoids new cross-layer signaling and therefore integrates with existing deployments. Second, by delegating parity generation to an off-the-shelf Reed-Solomon library, it keeps the coding machinery standard and well understood. The average coding rate of 0.94 shows the adaptation is conservative: overhead is added only when the estimated loss warrants it, so throughput cost remains modest while tail latency improves substantially. The comparison against No-FEC isolates the contribution of adaptive parity rather than scheduler behavior, supporting the claim that the approach is scheduler-agnostic. Limitations follow from the evaluation scope: the field trials cover three commercial LTE/5G paths at a 4.0 Mbps setting, so generalization to other rates, path counts, and radio technologies remains to be established. The independent-loss approximation used for parity selection may also be optimistic under correlated losses, though the measured 95th-percentile improvement suggests the scheme remains effective in practice. Overall, the work positions adaptive FEC as a deployable complement to coded multipath designs, lowering the barrier to practical adoption in vehicular cellular environments.

CS practitioners and researchers

Opening member content…