S. Salsano| SRv6 for the AI Backend| SRv6 Workshop — Netdev 0x1A, Rome 2026
University of Rome Tor Vergata CNIT
SRv6 Workshop @ Netdev 0x1A

SRv6 for the
AI Backend

Stefano Salsano1,2
1University of Rome Tor Vergata, Italy  ·  2CNIT, Italy
Netdev 0x1A — The Technical Conference on Linux Networking
SRv6 Workshop — 13 July 2026, Rome, Italy
C. Filsfils, P. Camarillo, A. Abdelsalam, A. Quinci, A. Tulumello, A. Mayer, P. Loreti, L. Bracciale, S. Salsano,
“Toward Deterministic Path Placement in AI Backends: A Practical SRv6-Based Architecture”, IEEE CNSM 2025
Part 1 — The problem

An AI backend is not the Internet

  • One owner, one fabric, known topology
  • Few, huge flows — not millions of small ones
  • Collectives are synchronous: all-reduce, all-to-all
  • The slowest flow sets the job completion time
  • GPUs idle while the network catches up
Consequence. Average throughput is irrelevant: the tail is what you pay for.
Spine 1 Spine 2 Leaf A Leaf B Leaf C GPU host GPU host GPU host all-reduce: everybody waits for the last byte
Part 1 — The problem

ECMP: spreading is not placement

  • The switch hashes the 5-tuple — it does not plan
  • Hash collisions: two elephants, one uplink
  • No entity owns the end-to-end path
  • On failure: wait for control-plane convergence
  • Rehashing moves flows you did not want to move
The gap. We spread flows statistically on a fabric we control exactly.
Spine 1 Spine 2 (idle) Leaf A Leaf C same hash bucket 2 × 400G on one 400G uplink flow 1 flow 2 half the fabric is idle, half the job is late
Part 2 — The fix

Source routing: the segment list is the path

1
Decide at the source. The host (or the NIC) picks the exact fabric path, per packet or per flowlet.
2
No hash, no collision. Placement is explicit: the fabric stops guessing and just executes the SID list.
3
Failures are a local decision. Switch the SID list at the host — no waiting for control-plane convergence.
And it is plain IPv6. SRv6 needs no new dataplane in the fabric — the switches already forward it.
Part 2 — The fix

Deterministic, congestion-aware placement

  • Telemetry in: link load, queue depth, failures
  • Path set out: a few disjoint SID lists per destination
  • The host selects, packet by packet
  • Reaction time = host RTT, not convergence time
Division of labour. The controller proposes the paths, the host disposes of the packets.
Controller / telemetry candidate SID lists Host / NIC — per-packet selection <S1, S3, D> <S2, S4, D> pick the path that is not congested Fabric: just forwards the SRv6 packet no new dataplane, no hash lottery
Part 3 — Linux

The Linux datapath

  • seg6 lwtunnel: encap, one static SID list per route
  • seg6local: the behaviors, on the receive side
  • The hard part: choosing the list per packet
  • Not a routing lookup — a policy decision
  • Needs feedback state next to the datapath
The kernel gap. seg6 can apply a path; nothing in it picks one.
# one path, statically bound to a route ip -6 route add fc00:d::/64 \ encap seg6 mode encap \ segs fc00:1::s1,fc00:3::s3 dev eth0 # what an AI backend actually needs for each packet: p = select_path(dst, telemetry, state) push_srh(p)
Where the selection lives
eBPF at the socket / TC egress — or XDP, or the NIC itself.
Part 3 — Linux

Doing it at line rate

Where it runsPer-packet choiceLine rate at 400G+
seg6 lwtunnel kernel, route-bound no — static list CPU-bound encap
eBPF / XDP kernel, programmable yes — maps hold the path set good, but CPU cycles per packet
NIC offload hardware yes, if the NIC exposes the SID list the only way to keep the GPUs fed
The ask to this room. A clean split: policy in eBPF, encap in the NIC — with a kernel API that does not fight either.
Part 4 — The ecosystem

It does not work alone

●
FRR. SID allocation and distribution — the host needs to know which segments exist.
●
SONiC. The switch side of the same story — see Ahmed's talk right after this one.
●
Telemetry. Load and failure signals, fast enough to be worth reacting to.
★
MRC. This architecture underpins the industry MRC protocol — main-track talk at this conference.

Conclusions

What we have

An SRv6 architecture that turns an AI fabric from statistically spread into deterministically placed — on plain IPv6, no new dataplane.

What we need

Kernel and NIC support for per-packet segment-list selection at line rate, and the telemetry loop to drive it.

Take-home. In a fabric you own completely, guessing the path is a choice — and a bad one.

C. Filsfils, P. Camarillo, A. Abdelsalam, A. Quinci, A. Tulumello, A. Mayer, P. Loreti, L. Bracciale, S. Salsano,
“Toward Deterministic Path Placement in AI Backends: A Practical SRv6-Based Architecture”, IEEE CNSM 2025
stefano.salsano@uniroma2.it