Hierarchy readers at tens of millions of signals: store scope id + leaf name, not a full path per signal
Profile of a waveform hierarchy reader on a 5M-signal, ~1M-scope file: single-threaded, ~3 heap allocations and ~480 bytes per name, cost per name doubling as paths got longer, because every variable stored its full hierarchical path string (often twice: once for display, once as a hash-map key). Linear projection to 50M signals at depth ~42 gave ~24 GB for the walk, ~90 GB for planning and >200 GB for conversion, far beyond a 32 GB box. Ranked fixes: (1) store scope id + leaf name per signal and build full paths on demand (bytes per name drop ~5x, memory stops growing with depth); (2) remove eager per-record text formatting, e.g. id-to-hex (~13% here); (3) pool small allocations; (4) compact sorted records instead of strings in later passes; (5) only then multi-thread the parse. Run a profile first (top functions, allocations by site, phase split) before choosing; the obvious guess, threading, ranked last. Keep output byte-identical as the gate for every step.