Allocator benchmark report

Thread scaling (sparse sweep)

mixed rigor - 3-block legacy coverage plus 40-repetition distribution bands. Legacy panels use 3 blocks per cell, median with min/max, and deliberately no confidence intervals or noise gating.

Worker counts are literal 1, 2, 3, 4, 6, 8; the runner allows 4 logical CPUs, so higher points are oversubscribed and describe contention rather than core scaling. splitmix64-v1 chain over (master seed, workload tag, worker-count cell, repetition/block, worker index); allocator identity, time and scheduling are absent; distribution size streams derive from the worker seed with the size-v1 domain. Scaling run 37637875091/1 measured mimalloc-pprof at source ee3623f6b7a485754a0a0aa3462d8cba43390e8b, which is not necessarily the commit above; metric key 0fa9e38615971bfa1f59193e92fb6ec8db5659f390d1ba8c7c5b52d53eb7c130.

Deterministic requested-size distributions

These are normal allocation requests, not pointer-alignment tests. Each line is the median of at least 40 paired repetitions. Every allocator shares one zero-based Y domain and ticks per metric, computed from every raw observation across both workloads and rounded upward; outliers are retained and never clipped. The P5 / P50 / P95 spread is in the table below. On every RSS chart the dashed live data (theoretical minimum) line is the least RSS any allocator could show: the process's own RSS before any worker started, plus the requested bytes the workload held at once (measured by a replay of the same stream, the smallest across every run of the cell). The table gives it as Floor MiB, and how far above it each allocator's median peak sits.

RSS after the work stops

The short-lived-thread stream (the large-class-ephemeral workload at 8 workers, replayed block for block from the same run), after which every worker thread is joined and the process stays alive and idle, like a server between requests. Each child reads its own /proc/self/statm at fixed times after the drain; peak RSS is the external peak. Released by is the first sample within 1.0 MiB of the 3 s sample, perf-ab's definition. The dashed live data (theoretical minimum) line is where an allocator that returned everything would land: nothing is live after the drain, so it is the process's RSS before any worker started. Median of 40 paired runs; the P5/P95 spread is in latest.json.

Pending panels contain no measured values.