# cgroup v2 lever runs

One directory per 3-minute run on a t3.micro, Amazon Linux 2023, kernel 6.18, no swap unless the run adds zram. Produced by ../../lab/cgroup-levers/driver.sh.

- collect.csv: 5-second samples (memory and io PSI, major faults, page-ins, file refaults, meminfo, swap used, oom_kill, cold python3 start time in ms).
- meta: start time, heap size, heap fill, final unit states.
- agent.log: the agent probe's pass times.
- oomd.log, dmesg.oom, zram.mm_stat: present when non-empty.
- summary.txt: medians after the first 30 s, from summarize.sh.

Batch A: r0, r1, r2, r3, r4, r5, r6, r7, r8. Batch B: r0c, r2b, r4c, r8b, r9c.
r5 and r6 used an earlier "half" fill (mrand48-based), so their heaps compress better than the current anon-hog half mode.
Four runs were dropped: three inherited a stale io.max from r2b, and one used a fill that was not actually incompressible.