Constraints for coded tunnels across long latency bottlenecks with ARQ-based congestion control

This paper considers capacity and delay constraints for coded tunnels across an erasure channel which occurs on shared Internet satellite links. Such links are long latency bottlenecks with a limited memory input queue which drops packets when it overflows. The latency delays ARQ ACK feedback to senders, making it difficult for them to tune their packet transmission rate. This can cause the input queue to oscillate between empty and overflow. Queue oscillation leaves the link underutilised during the empty phases and slows down large packet flows. Channel coding can in principle provide goodput improvement in this scenario by letting senders accelerate to higher packet rates before burst losses occur and by mitigating exponential backoff after losses. However, this is only possible if the codes preserve sufficient spare channel transmission rate for the improved goodput to expand into. We formulate rate and delay constraints that such block codes must meet. Using loss data obtained on a purpose-built simulator network, we show that such coding is feasible in a practical scenario and that partial unit memory (PUM) codes are particularly suitable for this task. In this context, we propose a part-systematic encoding for PUM codes, which performs slightly better than non-systematic encoding.