QFabric: Guarded Cross-Domain Task Recovery and Its Limits on a Linux-RTOS Board

QFabric. Guarded Cross-Domain Task Recovery and Its Limits on a Linux-RTOS Board. By Ayushman Bhattacharya and one other, from Pollinations AI and JIS University. Published on the first of November.

Some embedded boards pair a Linux application processor with a real-time microcontroller. The pairing invites an obvious move: when a task's timing starts to degrade on one side, shift it to the other. This work tests that idea and reports its limit. On heterogeneous boards where Linux does the dispatching, migration cannot isolate a task from Linux interference. Guards on recovery still matter, but mainly because they stop harmful moves from happening.

The vehicle is QFabric, a prototype that treats placement across domains as guarded recovery from a violated empirical soft real-time contract. A bounded task is compiled for both domains and monitored in windows with hysteresis. It is switched only after passing semantic and capacity admission at a quiescent epoch boundary, and the move is then committed or rolled back after a period of probation. Each of those steps is a guard: the monitoring windows carry hysteresis, the switch waits on admission of two kinds, and even a move that passes them all can still be rolled back once probation ends.

The board is an Arduino UNO Q, with a Cortex-A53 running Linux and a Cortex-M33 running Zephyr. The authors characterize its stock Bridge RPC interconnect and measure five workloads in both domains, under idle conditions, under CPU load, and under higher-priority SCHED_FIFO contention, with both ordinary and real-time callers. Every MCU invocation costs about nineteen milliseconds of dispatch, whatever the compute. The MCU runs the heaviest kernel thirty-four times slower. And because every call passes through Linux userspace, contention on Linux inflates the MCU path far more than the Linux path. At p ninety-nine, the MCU path reaches up to four hundred milliseconds, while the Linux path stays at seventeen milliseconds.

Across thirty cells of a four-node chain, Linux meets every contract with misses within zero point eight percent, and QFabric correctly never migrates. A controller without hysteresis migrates in seven cells, raising p ninety-five by as much as one hundred and seventy-seven times. Probation rolls back five of those moves.

The conclusion is stated for Linux-dispatched heterogeneous boards, and the measurements come from one board, five workloads, and the contention settings described. The point that generalizes within that scope is structural: if every call to the microcontroller has to traverse Linux userspace, the microcontroller cannot be a refuge from Linux trouble. The authors also note that this work supersedes an earlier artifact release, whose reported improvement came from measurement defects that they now identify.

The outcome inverts the usual pitch for such a system. The success case here is a controller that declines to act, and the measurable value of the guards lies in the harm they prevent. Pairing that with an open correction of their own earlier numbers, the authors make the limit itself the result.