Questions to ask a gateway vendor about what survives a partial failure

A central gateway rarely fails all at once. It fails in pieces, and the vendor's architecture decides which pieces take the others down with them. When the routing core, the hypervisor partitions and the Ethernet switch share silicon, every redundancy claim on the datasheet is really a claim about isolation. With limited lab time, your job is to ask questions that force the vendor to show where that isolation ends.

What still routes when the application partition hangs

Ask this first, and ask for the answer as a map of shared resources rather than a list of partitions. The classic failure mode in consolidated gateways is a non-safety partition, often the one running telematics or OTA services, that locks up while holding a shared resource: the Ethernet switch driver, a DMA channel, the interrupt controller or simply memory bandwidth. The hypervisor stays healthy, the watchdog keeps getting serviced, and safety-relevant CAN frames stop reaching the zonal controllers. Functional safety standards such as ISO 26262 call the required property freedom from interference, and the vendor should be able to show spatial and temporal evidence for it on every resource the safety path touches, not just on CPU cores. If the answer never gets past separate VMs, what you are buying is a common cause failure dressed up as partitioning.

How fast is failover, and measured from where

Every vendor quotes a switchover time, and the number means nothing until you know both endpoints. A figure such as 50 ms for restoring safety-relevant routing only fits your fault tolerant time interval if it runs from the instant of the injected fault to the first valid frame on the destination bus, measured on that bus. Many reports start the clock at detection and stop it when the backup task reports ready, which hides detection latency, network reconvergence and the time the receiving ECUs need to back out of their own timeout reactions. Ask for the measurement setup, the injection point and the worst case over repeated runs, not the mean. The risk this exposes is a failover that is fast on paper and still too slow for the function that depends on it.

Rollback that survives its own failure

Ask what happens when power drops while the inactive bank is being written, because a rollback path that depends on the same bootloader, key store or partition table the update is modifying is not a fallback, it is a second copy of the single point of failure.

A decision rule for limited bench time

You will not reproduce the vendor's validation campaign, and you should not try. Pick one fault the vendor did not choose, ideally a hang in the busiest non-safety partition under full network load, and ask them to inject it on your bench or on a remote rig you can watch live. My estimate is that this costs about a day of joint work per shortlisted supplier, which is cheap next to finding the problem at vehicle integration. The rule is simple: if the safety-relevant routing path shares any resource with a non-safety partition and the vendor cannot show interference evidence for that resource, score the path as a single point of failure, whatever the redundancy slide says. Even the vendors who pass leave you one open trade-off, since stronger isolation eats throughput and latency headroom that the next software release will want back.

Can a simulation replace fault injection on the bench?

For timing, only partly. A virtual ECU model can show that the failover logic is correct, but the interference you are worried about lives in caches, DMA arbitration and switch hardware, which most simulations abstract away. Accept simulation for breadth of coverage and insist on real silicon for the one fault that decides the score.

Back to AutomotiveGateway · Domains

← Back to AutomotiveGateway
Engineering & Architecture Brief

Questions to ask a gateway vendor about what survives a partial failure

A central gateway rarely fails all at once. It fails in pieces, and the vendor's architecture decides which pieces take the others down with them. When the routing core, the hypervisor partitions and the Ethernet switch share silicon, every redundancy claim on the datasheet is really a claim about isolation. With limited lab time, your job is to ask questions that force the vendor to show where that isolation ends.

What still routes when the application partition hangs

Ask this first, and ask for the answer as a map of shared resources rather than a list of partitions. The classic failure mode in consolidated gateways is a non-safety partition, often the one running telematics or OTA services, that locks up while holding a shared resource: the Ethernet switch driver, a DMA channel, the interrupt controller or simply memory bandwidth. The hypervisor stays healthy, the watchdog keeps getting serviced, and safety-relevant CAN frames stop reaching the zonal controllers. Functional safety standards such as ISO 26262 call the required property freedom from interference, and the vendor should be able to show spatial and temporal evidence for it on every resource the safety path touches, not just on CPU cores. If the answer never gets past separate VMs, what you are buying is a common cause failure dressed up as partitioning.

How fast is failover, and measured from where

Every vendor quotes a switchover time, and the number means nothing until you know both endpoints. A figure such as 50 ms for restoring safety-relevant routing only fits your fault tolerant time interval if it runs from the instant of the injected fault to the first valid frame on the destination bus, measured on that bus. Many reports start the clock at detection and stop it when the backup task reports ready, which hides detection latency, network reconvergence and the time the receiving ECUs need to back out of their own timeout reactions. Ask for the measurement setup, the injection point and the worst case over repeated runs, not the mean. The risk this exposes is a failover that is fast on paper and still too slow for the function that depends on it.

Rollback that survives its own failure

Ask what happens when power drops while the inactive bank is being written, because a rollback path that depends on the same bootloader, key store or partition table the update is modifying is not a fallback, it is a second copy of the single point of failure.

A decision rule for limited bench time

You will not reproduce the vendor's validation campaign, and you should not try. Pick one fault the vendor did not choose, ideally a hang in the busiest non-safety partition under full network load, and ask them to inject it on your bench or on a remote rig you can watch live. My estimate is that this costs about a day of joint work per shortlisted supplier, which is cheap next to finding the problem at vehicle integration. The rule is simple: if the safety-relevant routing path shares any resource with a non-safety partition and the vendor cannot show interference evidence for that resource, score the path as a single point of failure, whatever the redundancy slide says. Even the vendors who pass leave you one open trade-off, since stronger isolation eats throughput and latency headroom that the next software release will want back.

Technical FAQ: Can a simulation replace fault injection on the bench?

For timing, only partly. A virtual ECU model can show that the failover logic is correct, but the interference you are worried about lives in caches, DMA arbitration and switch hardware, which most simulations abstract away. Accept simulation for breadth of coverage and insist on real silicon for the one fault that decides the score.

Related Architecture Specifications (AUTOMOTIVEGATEWAY): automotivecompute.com · vehiclegateway.com · vehiclecompute.com

More Technical Articles on automotivegateway.com

Why Technical Domains Decide Authority in the AI Era →CAN Gateway Filtering, Routing and Diagnostics →Software-Defined Vehicle Reference Stack →ISAC and V2X Convergence for Cooperative Mobility →Centralized E/E Architecture: Consolidating the Vehicle →Zonal Architecture: Wiring, Power and the Ethernet Backbone →Automotive High-Performance Computing (HPC) →Automotive Ethernet and TSN: The Deterministic In-Vehicle Backbone →