MP stress campaign with coverage, all six boards (night 2026-11-17/18)
13 hours per board, rb-mp-stress at commit 515fc3e, one library per reset, 1,000,000 values per library and run. Prescaler 1-4 in rotation; at prescaler 3 and 4 half the runs use aimed interrupts and every run records where its interrupts landed (method: 2026-11-17-mp-coverage-prototype.md). The first campaign of this kind, and the last long one before version 1 (DECISIONS 2026-11-18).
Result
No errors in any library in its documented mode, in any run on any board. A run is valid when the control (an SPSC buffer used by two producers or two consumers) was caught in it.
| Board | Runs | Valid | Control caught, random interrupts | Control caught, aimed |
|---|---|---|---|---|
| nRF5340 DK | 64 | 53 | 37 of 48 | 16 of 16 |
| NUCLEO-G0B1RE | 116 | 105 | 77 of 88 | 28 of 28 |
| NUCLEO-G474RE | 218 | 206 | 152 of 164 | 54 of 54 |
| NUCLEO-H723ZG | 143 | 143 | 107 of 107 | 36 of 36 |
| SAM D20 Xplained Pro | 74 | 66 | 48 of 56 | 18 of 18 |
| Pico 2 | 235 | 177 | 119 of 177 | 58 of 58 |
| All | 850 | 750 | 540 of 640 | 210 of 210 |
No run failed (probe, serial port or build). Every run where the control was not caught had random interrupts at prescaler 2, 3 or 4. On the Pico 2 no random run at prescaler 3 or 4 caught it (0 of 58); at prescaler 1 and 2 all 119 did.
Coverage (prescaler 3 and 4, valid runs, the control left out)
Per library and mode: the share of interrupts that landed in the library's code, the median number of interrupts per instruction boundary where one can land, and how many of those boundaries were interrupted at least 10 times in the weakest run. Range over the libraries:
| Board | Interrupts | In library code | Median per boundary | Boundaries >= 10x, weakest run | Rows with every boundary >= 10x |
|---|---|---|---|---|---|
| nRF5340 DK | random | 2.0-8.7 % | 8-3,014 | 47-88 % | 0 of 13 |
| aimed | 8.7-32.3 % | 2,880-9,790 | 63-100 % | 1 of 13 | |
| NUCLEO-G0B1RE | random | 6.2-14.5 % | 33-2,744 | 55-89 % | 0 of 11 |
| aimed | 14.3-42.9 % | 704-15,730 | 84-100 % | 8 of 11 | |
| NUCLEO-G474RE | random | 2.3-11.1 % | 83-3,383 | 40-80 % | 0 of 13 |
| aimed | 4.2-27.8 % | 2,188-18,481 | 68-100 % | 4 of 13 | |
| NUCLEO-H723ZG | random | 8.8-20.0 % | 8-518 | 42-75 % | 0 of 13 |
| aimed | 9.0-26.1 % | 52-5,590 | 50-100 % | 1 of 13 | |
| SAM D20 Xplained Pro | random | 9.1-21.0 % | 96-6,530 | 50-91 % | 0 of 11 |
| aimed | 18.2-53.6 % | 433-12,289 | 85-100 % | 8 of 11 | |
| Pico 2 | random | no valid run | |||
| aimed | 1.9-24.5 % | 1,712-7,970 | 61-100 % | 2 of 13 |
Per-library tables: report.md in each campaign directory
(~/lab-runs/mp-campaign-20261117T1736*Z/ on lib-lab1), boundary lists in
coverage.txt.
Reading
- The aimed runs are the ones that test: they caught the control in every run on every board. With them the lowest median per boundary is 52 (heapless on the H723); on the other boards it is 433 or more. With random interrupts no library on any board had all its boundaries interrupted 10 times in every run.
- Random runs at the slow rates add little. On the Pico 2 they are never valid, so half of its prescaler 3/4 time (58 runs of about 3 minutes) gave nothing.
- The aim window is one value (160 cycles) for every library and board. The H723 gains least from aiming (not examined why), and the long lock-free operations (heapless, reserve/commit) keep boundaries below 10 on most boards. A window per library would close that; it is a method change and waits until after version 1.
- Boundaries in blocks that were never interrupted (heapless: about 90 boundaries on the Cortex-M4/M33/M7 boards, 41 on the Cortex-M0+ ones) are paths the test does not take, likely retries after a real collision. An interrupt test on one core does not reach them; model checking does.
For the publication re-run
No change is made here. A proposal for the owner, since it changes what a campaign covers: at prescaler 3 and 4 run aimed interrupts only (random runs there are often invalid and never on the Pico 2), and keep random interrupts at prescaler 1 and 2, where the control is caught reliably.