Joseph Doherty
d8a000a409
v2-ci / unit-tests (push) Has been cancelled
Profiled the remaining lmxopcua#789 misses on windev, a VMware VM, using per-thread CPU and context-switch accounting plus ETW CPU sampling. - One blocking hand-off costs ~26 us of latency and ~18 us of CPU there, about 10x macOS. - An idle evaluation makes four hand-offs in a row: host writer -> worker reader -> evaluation thread -> host reader -> dispatcher. - 48% of host CPU at 10k/s is in the kernel: the per-RunAsync NtWriteFile, thread-pool wakes, reads and waits. Spin-before-block was measured at every hand-off and is not kept. It meets 150 us idle p50 (118 us) only by keeping about two host cores and four worker cores busy for one stream. At 10k/s it reads +250 points. Harness: new "paced 10,000/s, batched" row. It uses the same rate and keys, with each 1 ms tick submitted as one RunBatch call, which is the production seams' shape. It runs after the original four phases, so their rows stay comparable. On windev it halves host CPU, to +36..+45 points against +71..+84 unbatched. That is still above 25. Design doc: - 12.8: the record. - 14 #15: the owner's decision on the Windows budgets. Recommended: measure on bare-metal Windows, then state the two budgets per platform, with the batched row as the gate. The alternative is a shared-memory transport. Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi