pq-all and trackmemory restore a seed ccache that already sits at its
cap, so every object a run compiles evicts a seeded one that is still
needed. Both reported a cache 100% full with over 200 cleanups and a
43-44% hit rate, and their configs took 3-5x their warm time. That left
pq-all shards finishing at 23-26 min against a 30 min limit: a healthy
shard passed with 5.6 min to spare, so any hiccup ahead of the build
took the job down. One did, twice, when the prebuilt deb image came up
short and the dependency step fell back to apt for 9.7 min - the only
two jobs out of 47 whose install ran past 2 min are the only two that
timed out. Size both caches to hold the seeded objects and a run's own
alongside them, and widen the limits so a slow install is survivable.
Raise the trackmemory limit as well, for a different reason: its
wolfEntropy configs have overrun it outright, once with six configs
done inside 8 min while all-wolfentropy ran on for another 32. That is
a stall rather than a slow run - 91 jobs have a 17.2 min median and a
17.9 min p90, with nothing between 18 min and the cap - so the higher
limit is headroom, not a fix. It costs nothing on a healthy run.
Recalibrate "minutes", which drives longest-first scheduling and shard
balancing. pq-all and fips-dev-no-post had drifted several-fold in both
directions: two frodokem cross-builds were declared at 4 min but take
0.3, so the pool kept scheduling them first as its longest work. pq-all
is calibrated from runs whose ccache actually hit, which the resized
cache makes the normal case; a value read off a cold run makes every
warm run warn instead. os-check-linux was already calibrated cold and
correct, so only its three genuine outliers change.
Both config lists are re-sorted to restore the documented descending
order, so most of their diff is a mechanical move rather than a value
change. Shard totals come out at 29.4/29.6 for pq-all and 76.9-77.7
across the four os-check shards.
Tobias Frauenschläger
9a788724d7