Random linear network coding for streams with unequally sized packets: Overhead reduction without zero-padded schemes

Abstract

State of the art Random Linear Network Coding (RLNC) schemes assume that data streams generate packets with equal sizes. This is an assumption that results in the highest efficiency gains for RLNC. A typical solution for managing unequal packet sizes is to zero-pad the smallest packets. However, the efficiency of this strategy depends heavily on the packet size distribution and can significantly curb the gains of network coding in practice. The work by Compta et al. showed that this strategy could introduce an overhead as high as 100% for Internet traffic. Our work advocates for an alternative encoding and decoding designs, focused on processing macro-symbols (composed by a number of symbols in the appropriate finite field) instead of full zero-padded packets. Our proposed schemes provide on-the-fly decoding strategies to manage heterogeneous packet sizes without the need for fragmentation or bundling of packets, which are strategies that require additional signaling. Performance is evaluated using CAIDA TCP packets and 4k video traces. Our results show that our mechanisms reduce significantly the padding overhead even for small field sizes. Finally, our strategies provide a natural trade-off between computational complexity and overhead due to the low density of the code when decoded at a macro-symbol level.