FreeBSD Guard Memory: mimalloc 3.2.8 (MI_SECURE=ON) Group Wins with About 70% Lower RSS

Full experiment timeline #

The same 32 original relays averaged 4.08 GiB in May 6–June 4 and 0.58 GiB in June 12–July 11: an 85.8% reduction, excluding the first seven days after the June restart. The later 64 additions are shown separately; the October 31 expansion to 200 relays falls after this frozen timeline’s midnight endpoint.

Supporting data #

Matched 30-day comparison #

Both matched windows are in the FreeBSD 15.0 phase, with base jemalloc 5.3.0 and requested mimalloc 2.1.7 / 3.2.8 builds. Each original relay contributed 120 readings to each window. Values are mean RSS in GiB per relay. Cohort labels describe intended configuration throughout.

Earlier window: spread and peaks #

May 6–June 4; mean, median, 95th percentile (p95), and maximum across sampled relay observations, in GiB. These are descriptive distributions, not confidence intervals or peaks between samples.

Full pre-upgrade low-memory interval #

Across June 5–October 15, the five configured cohorts averaged 556.9–578.4 MiB per relay. This FreeBSD 15.0 interval uses base jemalloc 5.3.0 and the same requested mimalloc 2.1.7 / 3.2.8 builds. It contains 531 collection times, including startup and three August restarts; it differs from the matched 30-day comparison above. Values below are MiB.

B’s lower sampled maximum (654.1 versus A’s 998.9 MiB) does not establish that tuning caused it or impose an upper bound on future memory use. Peaks between collections were not measured, and historical allocator activation remains unverified.

Later windows: intended cohorts only #

These values are retained for completeness, not as a ranking of active allocators. The earlier column is the FreeBSD 15.0 / jemalloc 5.3.0 phase; the later column is FreeBSD 15.1 / jemalloc 5.3.0. C–E retain the requested mimalloc 2.1.7 / 3.2.8 labels defined above. The first window has 120 readings per relay; the second has 29, ending at October 31 midnight. Values are mean RSS in MiB per relay.

The 64 October additions averaged 437.9 MiB in the second window. Including them in A would change the comparison’s population.

Complete original-relay comparison #

Relay A1–E6 are anonymized labels within allocator groups A–E. Each Guard-period mean uses all 390 readings from February 27–June 4; each non-Guard mean uses 120 readings from June 12–July 11. These columns have different observation periods. The phase is FreeBSD 15.0; configurations are identified explicitly in each row.

Requested settings #

For base jemalloc 5.3.0 (FreeBSD 15.0), later base jemalloc 5.3.0 (FreeBSD 15.1), B requested background_thread:true,dirty_decay_ms:1000,muzzy_decay_ms:1000,narenas:2 through MALLOC_CONF; background threads are unsupported in the base FreeBSD build. C requested mimalloc 2.1.7 with MI_SECURE=OFF. D requested mimalloc 3.2.8 with MI_SECURE=OFF; E requested mimalloc 3.2.8 with MI_SECURE=ON, and the same runtime options:

MIMALLOC_PURGE_DELAY=0
MIMALLOC_PURGE_DECOMMITS=1
MIMALLOC_ARENA_PURGE_MULT=1
MIMALLOC_ARENA_EAGER_COMMIT=0
MIMALLOC_PAGE_FULL_RETAIN=1

Why the wider memory evidence matters #

The sequence of interventions is more informative than any one RSS snapshot. These are overlapping reports from a small number of studies, not one independent experiment per article. Linux absolute-memory figures below use the corrected binary units described in those studies’ source analysis; the small September study is summarized by a unit-independent relative difference.

Earlier Linux context #

Our Ubuntu allocator study and Debian five-way study concern different hosts, versions, workloads, and durations. Their numbers should not be treated as matched operating-system effects.

These are single-snapshot group means. The old generators labeled binary GiB values as “GB.” Ubuntu’s 3.0.1 group later reached 6.43 GiB. Debian’s historical jemalloc/glibc membership lists overlap by three relays; the chart’s assignment order leaves 17 effective jemalloc members. Those ambiguities further limit cross-study comparisons.