Getting 1+ Tbps from an x86 server#

Branimir Rajtar, 5x9 Networks

• Why: The datapath-internals story of the year at this conference — what it actually takes to make commodity x86 forward at terabit rates, relevant to anyone building or evaluating software dataplanes (vBNG, vEPC, NFV).
• What: The evolution of a virtualized BNG from many small 4-vCPU forwarder VMs to a single "Big VM" forwarder, ending at 200 Mpps per CPU and 1.6 Tbps on one server (~600 W). Includes an honest comparison table: ASIC white-label switching (multi-terabit, low flexibility, ~30k subscribers) vs. x86 (~1 Tbps and improving, ~200k subscribers, high flexibility, C/DPDK/eBPF/P4 development).
• How: SR-IOV plus Intel DPDK to bypass host kernel and drivers; systematic optimization in waves — newer hardware (+30%), deep DPDK/BIOS/PCIe/NIC tuning (+30%), then vTune-guided code-level redesigns driven by CPU cache behavior and PCIe bottlenecks. The horizontal-scaling small-VM design was abandoned because per-instance overhead dominated. CUPS architecture (dashboard/controller/forwarder) throughout.
• Where applicable: vBNG/BRAS is the use case and the headline numbers are tied to their specific hardware generations; the methodology (measure, profile, redesign around cache/PCIe) transfers to any DPDK dataplane effort.