PDUMint: A Case for PDU-Aware, In-network Packet Layout Translation

Abstract

Traditional packetization is a passive process dictated by network constraints such as MTU, often remaining oblivious to application-defined data units. This mismatch forces receivers to perform costly reassembly and boundary detection, thereby wasting precious computation cycles. In this paper, we argue that packetization should instead be an active, adaptive element tailored to the receiver's processing architecture. Reconstructing packet boundaries to align with application-level units simplifies receiver-side processing and enables a novel model that leverages application logic at per-packet granularity. This approach unlocks significant optimization potential, including hardware offloading, page-aligned zero-copy data placement, and stateless intermediate processing— all while eliminating the overhead of packet buffering and manual reassembly. To make these benefits practical to deploy, we propose in-network packet layout translation. Preliminary evaluation of our prototype shows that in-network layout translation enables receiver-side optimizations that depend on packet layout, improving throughput over conventional packetization and achieves 91% of the throughput of specialized sender-side data-unit-aware packetization.