Measure the generated MCP transfer burst after API dispatch isolation #524

Open
opened 2026-09-30 14:20:55 +00:00 by kayg · 0 comments
Owner

The generated MCP transfer profile in #484 needs a comparable release baseline after API task isolation fixed a File move stack overflow.

Local debug results on the shared 8-core host:

Profile Before isolation p50 / p95 ms After isolation p50 / p95 ms
32 sequential 64 KiB responses 35.82 / 61.96 40.06 / 59.72
16 concurrent 1 MiB responses 579.92 / 983.15 1322.07 / 1679.72

The burst p95 increased by 70.85%. Load averages differ: before 23.77 / 26.39 / 29.65; after 26.65 / 24.74 / 26.71. Burst CPU fell from 232.62% to 136.30%; peak RSS fell from 291110912 to 281198592 bytes. These measurements do not prove a quiet-host regression.

docs/perf/baseline.json has a release request-storm baseline (p50 22.8 ms, p95 67.3 ms, CPU 56.49%, peak RSS 166264832 bytes), but no matching generated-transfer profile. Payload and concurrency differ.

Run bench/parity.sh with the current release binaries on the perf VM while holding /root/perf.lock. Record load inside the lock. Compare the same payloads and concurrency. Performance does not block the #484 merge.

Code: a434634b2 (dispatch isolation). Job head: 413eb4fad2.

The generated MCP transfer profile in #484 needs a comparable release baseline after API task isolation fixed a File move stack overflow. Local debug results on the shared 8-core host: | Profile | Before isolation p50 / p95 ms | After isolation p50 / p95 ms | | --- | --- | --- | | 32 sequential 64 KiB responses | 35.82 / 61.96 | 40.06 / 59.72 | | 16 concurrent 1 MiB responses | 579.92 / 983.15 | 1322.07 / 1679.72 | The burst p95 increased by 70.85%. Load averages differ: before 23.77 / 26.39 / 29.65; after 26.65 / 24.74 / 26.71. Burst CPU fell from 232.62% to 136.30%; peak RSS fell from 291110912 to 281198592 bytes. These measurements do not prove a quiet-host regression. `docs/perf/baseline.json` has a release request-storm baseline (p50 22.8 ms, p95 67.3 ms, CPU 56.49%, peak RSS 166264832 bytes), but no matching generated-transfer profile. Payload and concurrency differ. Run `bench/parity.sh` with the current release binaries on the perf VM while holding `/root/perf.lock`. Record load inside the lock. Compare the same payloads and concurrency. Performance does not block the #484 merge. Code: a434634b2 (dispatch isolation). Job head: 413eb4fad294d9eaf42c8f28fe99bb963500711b.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
kayg/calternal#524
No description provided.