The 5GC performance problem — Gbps throughput, microsecond latency

The Throughput Wall

A single UPF in a Tier-1 operator deployment aggregates traffic from tens of thousands of subscribers simultaneously, each pulling video, gaming, or browsing at tens to hundreds of megabits per second. Aggregate throughput per UPF node reaches hundreds of Gbps to multiple Tbps. A standard Linux kernel network stack — designed for general-purpose computing — cannot sustain this. Every packet through the stack costs microseconds of CPU time for sk_buff allocation, protocol parsing, routing lookups, and socket layer traversal. At 10 Gbps with 1500-byte frames the per-packet budget is roughly 1.2…