MP stress test: one library per reset, validated against the old way (2026-11-17)
Plan step 5 (lab/plans/2026-11-16-mp-bundle-and-open-areas.md). The
two-interrupt (MP) stress test now runs one library per reset over the
serial port, from a bundle that also holds the ISR stress test's images
(run.sh PER_LIB=1 IMAGES=combined; RP2350 SMP: IMAGES=per_lib). Here it
is compared with the old way (one image with every library, parameters over
the debug port), with the same seeds and parameters.
Setup. All six boards; MPSC and SPMC at timer prescaler 1 and 3,
300000 values per library, fixed seeds per case; RP2350 also SMP-MPSC and
SMP-SPMC (roles on two cores). 26 cases x 2 ways, 52 records:
~/lab-runs/mp-validate-20261117/ (lib-lab1). Firmware: commit 01351d3
(the records written before ~06:35 UTC name a stale 86d20e7ca721+dirty,
see SOURCE_COMMIT_NOTE.txt there; reruns name f844fa4..693b73b,
which change only run.sh and uart_host.py, not the firmware).
Result
| Same verdict | Different | |
|---|---|---|
| Libraries (every one, every case) | 161 of 161 | 0 |
| Control caught | 23 of 26 | 3 (all SPMC, prescaler 3) |
(Plus 8 library-cases not available on the core both ways: DNedic lockfree on Cortex-M0+, which lacks compare-and-swap.)
- Libraries. Every library has the same verdict both ways in every case. The only faults are the single-core designs (pxq MPSC/MPMC/SPMC, RIOT tsrb, wizard97) on two cores in SMP mode, found both ways, and not in their documented mode there (not counted). pxq reserve/commit starves in SMP-MPSC both ways (not counted). No library in its documented mode had an error either way.
- Load. Interrupts per value agree within 0.1 in every case (e.g. 15.5/15.4, 6.8/6.8, 1.3/1.3): one library per reset loads the board as the one image did.
- Control. At prescaler 3 in SPMC the second interrupt rarely lands inside a pop, and the control is caught or not by chance either way: old missed it on G474, H723, RP2350; new on G0B1, G474. Where it was caught only lost/duplicated counts show it (no order errors). Campaigns already count a run without a caught control as invalid.
- Control error counts where caught: same order both ways (e.g. H723 MPSC p1 4707/4569, SAM D20 5742/5617, G474 MPSC p3 143/926 at the low rate).
Found on the way (fixed)
probe-rs resethung right after a download that wrote flash, on four boards started together: MP run.sh now tries three times (f844fa4).- The old way could record a run with the build defaults as if with the parameters sent (a firmware started by hand ran past its wait): run.sh now compares the header with the parameters (f1ae43d, 1c2c357). The one record it affected was rerun.
- A selection written over the debug port is lost to the STM32H7 RAM ECC
write buffer unless more words follow (b2252e0), and a hanging
probe-rs writethere gave a traceback (693b73b).
Conclusion
The new way gives the same results as the old one. Records of the two ways
stay apart (new method: names -perlib, field per_lib). Switching the
MP campaigns to the new way is the owner's call (plan step 5).