Failing Over at Peak Load: Qualifying a Sub-THz to Optical Gateway Against Its Own Control Plane

A sub-THz to optical gateway that switches protection within budget on an idle bench can still drop an inter-carrier peering session when the same fault hits at peak load. The data path survives. The control plane does not, because its keepalives sit behind traffic that has just been forced onto half the capacity. A qualification plan that times the switchover and nothing else will pass a gateway that brings down a neighbour's routing table in production.

Why switchover time is the wrong headline

Protection behaviour is usually specified against transport protection recommendations of the ITU ring and linear protection family, and those set the familiar target of restoring traffic within 50 ms. That figure measures how fast frames start flowing on the surviving path. It says nothing about what those frames are queued behind. At high utilisation, a failed optical line card pushes its whole load onto a working path that was already carrying close to its share, and the egress queues toward the inter-carrier interface fill within milliseconds. If routing keepalives, BFD packets and the fronthaul timing messages share a scheduler class with bulk traffic, or sit in a priority class that a misconfigured policer caps, they wait in line with everything else. The peer counts missed hellos, declares the adjacency down and withdraws routes, so the switchover you just measured as a success ends in a route withdrawal at the other carrier.

The test that exposes it

Drive the gateway with a traffic mix shaped like the busy hour rather than a flat line-rate blaster: sub-THz access flows with their bursty, beam-dependent arrival pattern, bulk transit across the optical core, and live control-plane sessions to an emulated peer carrier that runs its real hold timers. Push aggregate load to the point where the surviving path alone would exceed its capacity after the fault, because that is the regime where queues actually build. Then pull a working optical path, physically or through a fibre-cut emulator, at a random phase relative to the traffic bursts, and repeat across many trials so you sample the worst alignment and not the average. Instrument the peer, not the device under test: timestamp every keepalive and timing packet as it arrives at the emulated neighbour, record the largest gap between consecutive arrivals during the recovery window, and log whether any adjacency changes state. My estimate is that a few hundred randomised cuts are needed before the tail of that gap distribution stabilises enough to judge.

The decision rule

If the largest keepalive inter-arrival gap at the peer during any switchover at peak load exceeds one third of the negotiated hold timer, the gateway fails qualification, regardless of how fast the data path recovered.

What the rule forces on the design

A third of the hold timer leaves margin for two more lost hellos from ordinary causes, which is the reason the threshold is not set at the timer itself. Gateways that fail it usually share one fault: the control and timing traffic was given priority on paper but routed through a shared buffer pool that bulk traffic exhausts first. The fix is a dedicated queue with reserved buffers and a scheduler that serves it strictly before anything else, which costs buffer memory the data path would otherwise use, and as a rough estimate it shaves a few percent off the burst absorption available to access traffic. That cost has to be argued against the alternative, a gateway whose failures stay local only while the network is quiet. Once the reserved queue is in, the next item on the plan is to rerun the same cuts with the peer's hold timer tightened to its most aggressive supported value, because the neighbouring carrier will eventually tune it down without asking you.

Why measure keepalive gaps at the emulated peer instead of reading counters on the gateway?

The gateway's counters show what it scheduled, not what arrived, and the adjacency is declared dead by the peer's view of arrival times. Measuring at the receiver catches queuing delay, policer drops and serialisation behind bulk frames that the sender's own statistics cannot see.

Back to 6GBridge · Domains

← Back to 6GBridge
Engineering & Architecture Brief

Failing Over at Peak Load: Qualifying a Sub-THz to Optical Gateway Against Its Own Control Plane

A sub-THz to optical gateway that switches protection within budget on an idle bench can still drop an inter-carrier peering session when the same fault hits at peak load. The data path survives. The control plane does not, because its keepalives sit behind traffic that has just been forced onto half the capacity. A qualification plan that times the switchover and nothing else will pass a gateway that brings down a neighbour's routing table in production.

Why switchover time is the wrong headline

Protection behaviour is usually specified against transport protection recommendations of the ITU ring and linear protection family, and those set the familiar target of restoring traffic within 50 ms. That figure measures how fast frames start flowing on the surviving path. It says nothing about what those frames are queued behind. At high utilisation, a failed optical line card pushes its whole load onto a working path that was already carrying close to its share, and the egress queues toward the inter-carrier interface fill within milliseconds. If routing keepalives, BFD packets and the fronthaul timing messages share a scheduler class with bulk traffic, or sit in a priority class that a misconfigured policer caps, they wait in line with everything else. The peer counts missed hellos, declares the adjacency down and withdraws routes, so the switchover you just measured as a success ends in a route withdrawal at the other carrier.

The test that exposes it

Drive the gateway with a traffic mix shaped like the busy hour rather than a flat line-rate blaster: sub-THz access flows with their bursty, beam-dependent arrival pattern, bulk transit across the optical core, and live control-plane sessions to an emulated peer carrier that runs its real hold timers. Push aggregate load to the point where the surviving path alone would exceed its capacity after the fault, because that is the regime where queues actually build. Then pull a working optical path, physically or through a fibre-cut emulator, at a random phase relative to the traffic bursts, and repeat across many trials so you sample the worst alignment and not the average. Instrument the peer, not the device under test: timestamp every keepalive and timing packet as it arrives at the emulated neighbour, record the largest gap between consecutive arrivals during the recovery window, and log whether any adjacency changes state. My estimate is that a few hundred randomised cuts are needed before the tail of that gap distribution stabilises enough to judge.

The decision rule

If the largest keepalive inter-arrival gap at the peer during any switchover at peak load exceeds one third of the negotiated hold timer, the gateway fails qualification, regardless of how fast the data path recovered.

What the rule forces on the design

A third of the hold timer leaves margin for two more lost hellos from ordinary causes, which is the reason the threshold is not set at the timer itself. Gateways that fail it usually share one fault: the control and timing traffic was given priority on paper but routed through a shared buffer pool that bulk traffic exhausts first. The fix is a dedicated queue with reserved buffers and a scheduler that serves it strictly before anything else, which costs buffer memory the data path would otherwise use, and as a rough estimate it shaves a few percent off the burst absorption available to access traffic. That cost has to be argued against the alternative, a gateway whose failures stay local only while the network is quiet. Once the reserved queue is in, the next item on the plan is to rerun the same cuts with the peer's hold timer tightened to its most aggressive supported value, because the neighbouring carrier will eventually tune it down without asking you.

Technical FAQ: Why measure keepalive gaps at the emulated peer instead of reading counters on the gateway?

The gateway's counters show what it scheduled, not what arrived, and the adjacency is declared dead by the peer's view of arrival times. Measuring at the receiver catches queuing delay, policer drops and serialisation behind bulk frames that the sender's own statistics cannot see.

Related Architecture Specifications (6G): gateway6g.com · 6gboard.com · 6gpqc.com

More Technical Articles on 6gbridge.com

Benchmarking End-to-End Latency and Packet Loss in 6G Multi-Carrier Border Gateways →Quantum-Safe Inter-Carrier Signaling: Deploying ML-KEM and State-Based Cryptographic Handshakes at Border Points →