Engineering Internship – Networking Deep Tech: DPDK, Linux and Fast Packet Processing on Embedded Hardware - Paris - H/F
iliad-free Engineering Internship – Networking Deep Tech: DPDK, Linux and Fast Packet Processing on Embedded Hardware - Paris - H/F
Description
INTERNSHIP SUBJECT :
You will design, implement and measure a dataplane packet-processing function on one of our embedded targets, from the algorithm to the numbers on real hardware, and push the result upstream.
The focus is the dataplane and its coupling with Linux. Packets arrive on a NIC or a switch port, are classified, looked up, possibly tunnelled or encrypted, scheduled and forwarded, while routes, neighbours and interfaces come from the management and control plane. You will optimize each design option (run-to-completion, pipelines, polling vs interrupts with power budget constraints, kernel versus bypass, lookup structures, etc.).
Possible topics, adjusted to your profile:
- IPv4/IPv6, SRv6 routing
- A DPDK gateway
- XDP/AF_XDP or DPDK for a given workload
- Low-power dataplanes: power-aware polling, interrupt-mode RX and their effect on p99 latency
- Tunnel and IPsec (VXLAN, GRE, GTP-U, ESP) with hardware crypto offload
- Board bring-up: PHY drivers, phylink, DSA/switchdev and TC flower offload
MISSIONS :
The work runs from the design to measured, optimised and upstreamed code.
- Learn the stack: DPDK internals (PMDs, mempools, mbufs, rings), Linux kernel networking, our boards, and the state of the art on fast packet processing
- Build a reproducible test bench with TRex, pktgen or testpmd, following the RFC 2544 methodology, and establish the baseline numbers
- Design and implement the function in C, on DPDK, with GRout or in the kernel, with the data structures and algorithms the hardware budget allows
- Measure and optimize: profile with perf, ftrace, bpftrace and PMU counters; meet the Mpps, p99 latency, jitter and watts-per-Gbps targets
- Upstream: post patches on the DPDK, GRout, netdev, U-Boot or Buildroot lists and follow the reviews until they are merged
- Engineering practice: design doc, code reviews, final report and presentation. A talk at an open-source conference is possible depending on results
COMPUTER SCIENCE AND NETWORKING REQUIRED :
A strong systems and networking background is the core requirement: these functions are algorithms and data structures running against a per-packet budget of a few hundred nanoseconds. C, computer architecture and L2/L3 networking are mandatory, the other areas are a strong plus and can be deepened during the internship.
Computer science
- Computer architecture: Key concepts: caches and cache lines, NUMA, TLB and hugepages, DMA, PCIe, memory ordering, prefetching, branch prediction · Used for: per-packet budgets, mbuf and mempool layout, zero-copy I/O
- Operating systems, Linux: Key concepts: kernel and user space, interrupts versus polling, scheduling and CPU isolation, IOMMU and VFIO, device drivers · Used for: kernel bypass, exception paths, board bring-up
- Algorithms and data structures: Key concepts: hash tables and cuckoo hashing, tries and DIR-24-8, lock-free rings, bloom filters, batching and vectorisation · Used for: FIB lookups, flow classification, pipelines
- Concurrency: Key concepts: lock-free and wait-free structures, RCU, memory barriers, run-to-completion versus pipeline models · Used for: multi-core dataplanes, control-plane to dataplane synchronisation
- Systems programming in C for Linux: Key concepts: memory layout and alignment, undefined behaviour, SIMD intrinsics (NEON, SSE/AVX), cross-compilation · Used for: every line of the dataplane and the drivers
- Performance engineering: Key concepts: RFC 2544 methodology, PMU counters, profiling, latency percentiles and jitter, benchmarking pitfalls · Used for: proving every gain with numbers
- Embedded software: Key concepts: boot chain and U-Boot, device tree/DTS, Zephyr, Buildroot and Yocto, debugging without a console (JTAG, ftrace, crash dumps) · Used for: bring-up and maintenance on our boards
Networking
- Layer 2: Key concepts: Ethernet, VLAN and QinQ, bridging, LACP, PHY and MAC, DSA and switchdev · Used for: board networking, switch offload
- Layer 3: Key concepts: IPv4 and IPv6, ARP and ND, routing and FIB, ECMP, MTU and fragmentation, tunnels (VXLAN, GRE, GTP-U) · Used for: forwarding paths
- Routing protocols: Key concepts: OSPF and IS-IS, BGP, SRv6, netlink, FRR and BIRD, shortest paths and loop-free alternates · Used for: control-plane integration, convergence and fast reroute
- Quality of service: Key concepts: schedulers, token buckets, shaping, active queue management, DSCP · Used for: QoS pipelines
- Security: Key concepts: IPsec and ESP, crypto offload, ACLs, resilience to packet floods · Used for: secure dataplanes
TECHNICAL SKILLS
You should be ready to deep dive and learn one of the dataplane targets.
Then of course, you’ll practice:
- C on Linux, with attention to memory layout and performance
- Linux tooling: git, cross-compilation, perf, gdb
- Python for test automation and tooling For the dataplane targets:
- DPDK: PMDs, mempools and mbufs, rings and lcores, eventdev, graph library, rte_lpm and rte_fib
- Linux kernel: netdev and PHY drivers, XDP and AF_XDP, netlink, switchdev and TC offload
Background, but not mandatory:
- Arm64 SoCs, SmartNICs and DPUs (NVIDIA BlueField, Marvell OCTEON, NXP DPAA2)
- FRR or BIRD, VPP, OVS-DPDK, GRout
- Rust for tooling, CI with the DPDK Test Suite or LAVA
- A first public open-source contribution, on any project. We will read it
Le groupe Iliad s’engage pleinement en faveur de ses collaborateur·rice·s, à travers une politique stricte de non-discrimination et de lutte contre le harcèlement. Toutes les candidatures sont les bienvenues, quels que soient le parcours, l’âge, l’origine, le genre ou l’orientation sexuelle.
Skills
Sign in to read and join the discussion.