pub fn gauntlet_presets() -> Vec<TopoPreset>Expand description
Topology presets used by gauntlet mode.
Covers topologies from 4cpu-1llc-nosmt (4 CPUs) up through the
252cpu-14llc-smt / 252cpu-14llc-nosmt presets (252 CPUs,
near the KVM vCPU limit), plus 192cpu-11llc-smt (192 CPUs across 11
NON-uniform LLCs), spanning SMT, non-SMT (-nosmt), and multi-NUMA
(2numa-*, 4numa-*) families. The stock presets are built from
the defs / numa_defs tuple tables; 192cpu-11llc-smt is appended
explicitly because it carries per-LLC core counts and a forced CPU
budget the uniform tuple shape cannot express.
192cpu-11llc-smt targets schedulers whose per-LLC math assumes
equal-sized caches: ten LLCs of 18 logical CPUs and one of 12
(llc_cores = [9;10] + [6], packing width 9). It is VERIFIER-ONLY —
the gauntlet execution path reconstructs topology from a uniform
TopoOverride that cannot express llc_cores, so
for_each_gauntlet_variant skips non-uniform presets. The verifier
reaches it via
crate::test_support::TopologyConstraints::accepts_verifier, which
ignores default caps and imposes NO host-size bound. Its
forced_cpu_budget = Some(96) guarantees continuous overcommit (192
vCPUs time-slicing over <=96 host CPUs) every time the battery runs.
The full set is returned unconditionally; the only filter applied
here is the aarch64 retain below, which drops SMT presets
(threads_per_core > 1) because ARM64 CPUs have no SMT (the
non-SMT scale presets keep ARM64’s topology scale
coverage). The multi-NUMA presets are not filtered here: the
default crate::test_support::TopologyConstraints
(max_numa_nodes: Some(1)) excludes them at test-selection time
(via accepts / accepts_no_perf_mode) unless a test raises
the bound.