The heartbeat's one input walked all 335,544 Stadium cells every tick, about 15 ms on aarch64 TCG, longer than a tick. At fast pacing it took the CPU, and a birth missed the session's 300 s on CI. That was my 11-17 change. Captain Bob ruled: sample less often, scaled by the cells, measured first, as a slice each tick; the slope inference later. - stadium_heat_walk_step() walks STADIUM_HEAT_SLICE_CELLS cells per tick (Kconfig, 5500, about 0.25 ms on aarch64 TCG). - A completed walk is one sample. - The heartbeat's load takes a sample only when a walk is done. Test first: test_stadium_node (21 checks) and test_heartbeat (13). Both failed to build before. All kernel host tests pass. Four aarch64 boots in a row pass 18/18 each, BIRTH included. The evidence is in logs/20261119-heat-slice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
31 lines
1.7 KiB
Plaintext
31 lines
1.7 KiB
Plaintext
The heartbeat's load, walked a slice each tick (2026-11-19).
|
|
|
|
The cause of CI's aarch64 BIRTH stall (every head since 58d0e887):
|
|
- stadium_resident_heat_total(), the heartbeat's one input (11-17, "ONE adaptive heartbeat"), walked
|
|
every Stadium cell on every serviced tick;
|
|
- aarch64: 335,544 cells. One walk measured ~15 ms on TCG (walk-timing-session.txt): 925k-987k
|
|
ticks of the 62.5 MHz counter, against 625k per 10 ms base tick. That is 1.5 base ticks, and
|
|
about 6 of the fastest (2.5 ms);
|
|
- so at fast pacing the walk took the CPU, and a birth did not finish within the session's 300 s.
|
|
|
|
Ruled (Captain Bob): "#2 but with a multiplier and maybe a slope inference"; "Scaled by the cells";
|
|
"Measure first"; "A slice each tick"; the slope inference "Later".
|
|
|
|
Built:
|
|
- stadium_heat_walk_step(): walks the next STADIUM_HEAT_SLICE_CELLS cells (Kconfig, 5500: about
|
|
0.25 ms on aarch64 TCG, a tenth of the fastest tick). A walk done gives one sample, so the ticks
|
|
per sample grow with the cells;
|
|
- the heartbeat's load type takes a sample only when a walk completes (timer.h, heartbeat.c);
|
|
- sk_v4.c gives the heartbeat the sliced walk.
|
|
|
|
Test first (tests-before.log: neither test built; tests-after.log):
|
|
- test_stadium_node, 21 checks: the walk in slices of 7 equals the full total, finishes after
|
|
ceil(cells/7), and starts again;
|
|
- test_heartbeat, 13 checks: a three-slice source gives a sample every third tick, and the slope
|
|
provider runs after HEARTBEAT_PACE_EVERY samples.
|
|
|
|
kernel-tests.log: all kernel host tests pass. build-aarch64.log: clean.
|
|
|
|
Four aarch64 boots in a row with the fix (run1-4.txt): 18 of 18 checks each, 453-458 s, BIRTH and
|
|
everything after it. The same loop caught the stall on its first boot before.
|