Rethinking Message Buffer Structures for 100 Mpps Cloud-native Network Functions

Extremely performant cloud-native network functions (CNFs) are created from mastering the effect of message (packet) buffer structures on hardware behaviors. Such CNFs are key enablers of in-situ 5G and 6G networks especially in network slicing, thanks to distinctive flexibility of software. Our previous work revealed that DPDK-accelerated CNFs leaved room for further optimizing the use of computing resources, and the efficiency of basic packet forwarding increased manyfold (over 150 Mpps) in ideal situation. However, metadata areas in message buffers had to be eliminated to achieve the ideal performance, resulting in impractical CNFs. In this paper, we revisit message buffer structures, and propose a two-phased approach, sizing and decoupling-and-aggregating of metadata areas, to remove this restriction. The experimental result showed that the efficiency of packet forwarding was enhanced to over 120 Mpps (per CPU core) when the size of metadata area was 64 bytes or less.