Modern networked systems are moving from hardware into flexible, programmable platforms. Among the key enablers of this shift we list the Intel DPDK and the eBPF ecosystem in conjunction with the eXpress Data Path (XDP), both allowing packet processing at very high rates.
As the classic per-packet processing model used by the usual networking stack can become a bottleneck (especially with pressure on CPU pipelines and the number of instructions per packet required), new designs require treating packets in batches, data prefetching, vectorized operations and better instruction-level parallelism. In this talk, we will show how kernel bottlenecks impact line-rate packet processing and how modern approaches attempt to overcome them, offering an accessible introduction to profiling, optimization, and programmable data paths in virtualized networking environments.
These ideas illustrate how rethinking execution models can significantly improve packet-processing efficiency. This talk provides an accessible introduction to these concepts, focusing on how XDP fits into virtualized networking, why profiling matters, and how batching-aware techniques can unlock new optimization opportunities for modern programmable data paths.
Bio: Vladimiro is a second-year PhD student at Sapienza University of Rome. At the beginning of his PhD, he focused on studying network functions and their performance. Together with his advisor, Salvatore Pontarelli, and his colleagues, he developed a profiler for eBPF applications that introduces only minimal overhead on the program being analyzed. He is now working on improving the performance of eBPF-based networking applications by bringing a batching model to eBPF.