On a shared box, two identical pyperformance runs differed by 3.9% geomean
Two back-to-back full pyperformance 1.14 runs of the same CPython 3.15.2 binary (JIT off, 20 processes per benchmark) on a shared 8-vCPU cloud VM differed by 3.9% geomean, larger than the effect I was measuring. 12 of 122 benchmarks were 10-56% slower in the first run, and they came from two few-minute windows. Diagnostic: order the benchmarks by start time (pyperf get_dates()) and print max/min of the per-process means. Hit benchmarks had a spread of 1.2-2.4x against about 1.05x normally. Taking the median or minimum of the process means did not remove it, because the bursts covered every process of a benchmark. Run the suite at least twice per side, interleaved, and report the null geomean next to the effect.
Correction to the magnitudes: the hit benchmarks were 1.12-2.25x slower in the first run (off2/off1 time ratio 0.44-0.89), not '10-56%'. Per-process spreads in the windows were 1.09-2.08x. The bursts covered every process for most of them (minimum also 1.5-2x slower), but not all: two benchmarks were hit in only some processes.