Leveraging Network Heterogeneity Weaknesses for Enhanced Streaming Communications

Streaming communications over erasure channels face fundamental challenges in balancing throughput, delay, and channel/spectrum utilization under delayed feedback and unknown channel models. Traditional solutions that maximize throughput struggle to effectively manage these tradeoffs, as they are typically aimed at the large regime, disregarding delay and utilization constraints.
In this talk, we demonstrate how network weaknesses can be transformed into opportunities for optimizing these tradeoffs in various scenarios. Our solutions build upon adaptive and causal random linear network coding (AC-RLNC) protocol, which schedules forward-error-correction (FEC) transmissions based on erasure rate estimation during the last round-trip-time (RTT) period. For single-hop streaming, we introduce DeepNP, a deep learning-based noise prediction model that estimates erasure rates using delayed feedback. While individual erasure events are difficult to detect, we facilitate prediction by focusing on overall erasure rates, as required by AC-RLNC and other streaming coding solutions. We further enhance prediction accuracy by incorporating SNR data from lower network layers, taking advantage of its continuous nature compared to binary acknowledgment feedback. This leads to a cross-layer scheme that balances coding rates at the physical layer with erasure rates at the transport layer, both determined by the modulation coding scheme. For a multi-hop system, we introduce Blank-Space AC-RLNC (BS), a novel solution designed to mitigate the trade-off between throughput, delay, and efficiency. BS leverages the network’s bottleneck constraints to introduce idle periods without degrading throughput-delay performance. Specifically, each node can implement Network-AC-RLNC (NET), a light-computational re-encoding algorithm that adaptively adjusts FEC rates and identifies ineffective transmission periods. NET incorporates two distinct suspension mechanisms: 1) Blank Space Period, which accounts for the bottleneck from each node to the destination, and 2) No-New No-FEC approach, which adapts to data availability. By simulations, we demonstrate that DeepNP significantly improves the throughput-delay tradeoff required for ultra-low latency communications in NeXt G and that BS achieves resource efficiency in multi-hop networks without compromising performance.

M.Sc. student under the supervision of Prof. Alejandro Cohen.