Skip to content
5G/6G AcademyEst. 2009
Join free
All Courses
Level 32

5G Performance Engineering — DPDK, eBPF, and Line-Rate Telecom

Engineer 5G core network functions for line-rate throughput and microsecond latency. Master profiling with perf and flamegraphs; understand where time goes in kernel networking; write user-space packet processors with DPDK; choose between DPDK, VPP, and AF_XDP; use eBPF for non-intrusive observability and XDP packet processing in the kernel; tune NUMA, CPU pinning, hugepages, and SR-IOV for production UPF; plan capacity at telecom scale. Pairs with Level 31 (NF Software Engineering) to complete the Telecom Software Engineer career path.

20 interactive0 simulation4 labs4 quizzes
0/28 lessons

Open 5G Performance Engineering — DPDK, eBPF, and Line-Rate Telecom free for 7 days

A new account opens this level, every other level and every certification course for 7 days — no card.

Join for free

Performance Foundations

0/7

Why 5GC performance matters, where time goes inside a kernel-based NF, the latency and throughput budgets that constrain UPF and SBI traffic, profiling fundamentals with perf and flamegraphs, and synthetic load generation for 5GC components.

The 5GC performance problem — Gbps throughput, microsecond latency
Why a UPF needs to handle tens of Gbps per node and why control-plane NFs care about sub-millisecond SBI latency.
interactive3 min
Where time goes — kernel networking, syscalls, context switches
A standard Linux network stack costs microseconds per packet — the cumulative drag that drives 5GC to kernel bypass.
interactive3 min
Performance budgets — RAN, control plane, user plane
Real 3GPP latency budgets translated into engineering constraints per NF and per-hop budgets.
interactive3 min
Profiling 101 — perf, flamegraphs, pprof
How to find the hot 5% of code burning 95% of CPU using standard Linux profiling tools and Go pprof.
interactive3 min
Synthetic load — iperf, k6, traffic generators for 5GC
How to generate realistic load for SBI control-plane and UPF data-plane components without real RAN.
interactive3 min
Lab: Profiling a misbehaving SBI handler
Walk through identifying a JSON-marshaling hotspot in an AMF handler using pprof and a flamegraph.
lab5 min
Quiz: Performance Foundations
Test your understanding of where time goes, performance budgets, and profiling tools.
quiz5 min

DPDK and User-Space Networking

0/7

What DPDK is and why telecom UPFs depend on it, DPDK architecture (PMDs, mempools, rings, hugepages), writing a minimal DPDK app, applying DPDK to UPF packet detection and forwarding, and how VPP and AF_XDP compare as alternatives.

What DPDK is and why telecom uses it
Intel DPDK explained — bypass the kernel network stack, poll the NIC directly from user space, achieve line-rate packet processing.
interactive3 min
DPDK architecture — PMDs, mempools, rings, hugepages
The five building blocks of every DPDK application and why each is essential for line-rate performance.
interactive3 min
Writing a minimal DPDK app
A 100-line C program that reads packets from a NIC, applies a filter, and writes them back — the smallest possible UPF.
interactive3 min
DPDK for UPF — packet detection rules at line rate
How a production UPF uses DPDK to apply PFCP-defined packet detection rules and QoS at millions of packets per second.
interactive3 min
VPP (Vector Packet Processing) — when to choose over DPDK
FD.io VPP processes batches of packets through a graph of nodes, often outperforming hand-rolled DPDK on complex workloads.
interactive3 min
Lab: DPDK testpmd packet-rate measurement
Walk through using DPDK's testpmd to measure the maximum packet rate of a server and identify the bottleneck.
lab5 min
Quiz: DPDK and User-Space Networking
Test your understanding of DPDK architecture, application patterns, and VPP alternatives.
quiz5 min

eBPF for Telecom

0/7

What eBPF is and why it transformed Linux observability, using eBPF for non-intrusive tracing of NF code paths, XDP for kernel-level packet processing, AF_XDP as a lower-friction alternative to DPDK, and how Cilium and similar tools deploy eBPF in production 5GC.

What eBPF is — running programs in the kernel safely
Extended Berkeley Packet Filter as a safe VM in the Linux kernel — and why this changed everything for observability.
interactive3 min
eBPF for observability — tracing without modifying NF code
Attaching eBPF probes to syscalls and kernel functions to measure NF behavior without source-code changes.
interactive3 min
XDP — eBPF-based packet processing in the kernel
eXpress Data Path: drop or redirect packets at the lowest possible kernel point, before sk_buff allocation.
interactive3 min
AF_XDP — kernel-bypass without DPDK's complexity
Combine kernel-managed NIC drivers with user-space packet rings for DPDK-like performance and standard Linux tools.
interactive3 min
eBPF in production 5GC — Cilium, observability stacks
How real 5GC deployments use Cilium for K8s networking and eBPF-based tracing for production NF observability.
interactive3 min
Lab: Tracing NGAP latency with bpftrace
Write a small bpftrace one-liner that measures AMF NGAP message handling latency without touching AMF source code.
lab5 min
Quiz: eBPF for Telecom
Test your understanding of eBPF, XDP, AF_XDP, and production eBPF tooling for 5GC.
quiz5 min

Production Tuning

0/7

NUMA topology and CPU pinning for UPF, hugepages configuration, SR-IOV for virtualized line-rate access, network interrupt tuning and RSS, and capacity planning for production 5GC at scale.

NUMA and CPU pinning for UPF
Why a UPF that crosses NUMA boundaries can lose half its throughput, and how to pin worker threads correctly.
interactive3 min
Hugepages — why and how
2 MB or 1 GB pages eliminate TLB misses for DPDK packet buffers — small change, dramatic effect on cache behavior.
interactive3 min
SR-IOV and Virtual Functions
Giving a UPF container direct hardware access to NIC virtual functions, bypassing the hypervisor and getting line rate in VMs and containers.
interactive3 min
Network interrupts, RSS, and kernel-bypass trade-offs
Receive Side Scaling, interrupt affinity, and the moment when polling beats interrupts for high-rate traffic.
interactive3 min
Capacity planning and load forecasting for 5GC
Real-world rules of thumb for sizing AMF, SMF, UPF, and NRF given expected subscriber and session counts.
interactive3 min
Lab: Tuning a production UPF for line rate
Walk through the full tuning checklist — NUMA, hugepages, SR-IOV, IRQ affinity, kernel config — that takes a UPF from 1 Gbps to 40 Gbps.
lab5 min
Quiz: Production Tuning — Level Gate
Final level gate covering NUMA, hugepages, SR-IOV, interrupt tuning, and 5GC capacity planning.
quiz5 min