Repository navigation
(perf): v0.0.96 idles at about twice v0.0.94's main-process CPU #691
Description
Observed (controlled benchmark, 17/10, one instance at a time, all runs under heavy machine load)
Main-process idle CPU, in % of one core, median of 3 runs per cell, alternating v0.0.96 (installer from aeabd72, same tree as 22d2fc7) and v0.0.94 on the same fixture:
| trace on | trace off | |
|---|---|---|
| v0.0.96 | 0.93 | 1.09 |
| v0.0.94 | 0.44 | 0.43 |
All 6 v0.0.96 runs (0.66–1.77 %) are above all 6 v0.0.94 runs (0.39–0.52 %). Every run was taken while NordVPN used ~45 % and Claude sessions ~30 % of the machine, so the gap is plausible, not confirmed. Data: D:\sb-qual-646\q096b\ and rca\out\q9* on the workstation.
#690 already removed the trace-only cost of the main-loop monitor (3–6 % → ~0.9 %). The remaining gap is the same with the trace on and off.
To do
- Re-measure on a quiet machine to confirm the gap.
- If confirmed, bisect by toggling one change at a time in scratch builds. Candidates: the always-on 1 s monitor tick, the writer worker ((index): move every session index write to a writer worker (#616) #646), the projects worker ((perf): build the projects list off the main process and load it once per reload (#616) #619), the descriptor/process-table checks ((agents): keep a background session hosted while it waits between turns (#682) #683).
Decided by Jean-Baptiste in person, 2026-10-17: v0.0.96 ships with this gap; the cause is tracked here.
Activity
devsuitup commentedon Nov 3, 2026
Decided by Jean-Baptiste in person, 2026-11-03: one quiet one-hour slot (VPN, Firefox and local sessions off) to take the #616, #674 and #691 measurements together.
devsuitup commentedon Nov 3, 2026
Quiet-slot measurement 2026-11-03 (VPN/Firefox off, machine load median 15.5 %), source runs from worktrees, Electron 41.0.3, fixture of 954 sessions, throwaway HOME. Comparable between versions, not with the earlier packaged absolute numbers.
Idle CPU, 300 s after scan + 60 s settle, trace off, 2 runs each (% of one core): main process — main 259011f 0.25 / 0.255, v0.0.94 0.052 / 0.068 (ratio ~4×); whole process tree — main 0.31–0.35, v0.0.94 0.14–0.17. The relative gap is real; the absolute cost is about a quarter of one percent of one core. Raw: .work-files/perf-slot/out-20261103-110033/summary.md (local)