Graviton (m7g.4xlarge) validation of thread-per-core serving before release

[membphis]

Split out of #891 (thread-per-core workers), which names this run as the release gate.

Every published number for the thread-per-core mode comes from a 12-vCPU x86 VM, where virtualized context-switch cost amplifies the win. The gain should be smaller on bare metal, and ARM is the deployment target that motivated the work.

Gate: run the #891 A/B (both proxy.thread_per_core legs, c=16/32/128 and the 10ms-TTFT high-concurrency point, fail=0) on an m7g.4xlarge against the merged branch.

If thread-per-core does not win there, flipping the Linux platform default back is a one-line change with no other code impact — that rollback decision is what this issue exists to make deliberately.

[membphis]

Validation ran on an m7g.4xlarge (16-vCPU Graviton3, Ubuntu 24.04, kernel 6.17), aisix built natively (release) from the PR branch HEAD b31de20. Same method as the #891 table: otb loadgen, local mock upstream, 25s windows, warmup first, fail=0 everywhere, CPU from /proc/<pid>/stat. Rig floor (mock direct, c=128): 475,521 rps.

4-worker A/B (gateway pinned 0-3, mock 4-7, loadgen 8-15 — same shape as the PR table):

c	thread-per-core	work-stealing	delta
16	26,489 rps · 400% CPU · p99 0.89ms	20,760 rps · 391% CPU · p99 1.21ms	+27.6%
32	27,990 rps · 400% CPU · p99 1.91ms	20,879 rps · 391% CPU · p99 2.39ms	+34.1%
128	29,533 rps · 400% CPU · p99 5.39ms	20,964 rps · 391% CPU · p99 10.07ms	+40.9%
768, 10ms TTFT mock	27,114 rps · 399% CPU · p99 33.6ms	19,215 rps · 388% CPU · p99 63.3ms	+41.1%

c=128 repeated four times: 29.5k / 29.6k / 29.8k / 29.5k against 21.0k / 20.8k / 20.9k / 21.0k. CPU per request ~187µs → ~135µs.

8-worker A/B (gateway pinned 0-7 — the wider-machine trend this gate worried about):

c	thread-per-core	work-stealing	delta
32	49,701 rps · 790% CPU · p99 1.12ms	38,605 rps · 778% CPU · p99 1.29ms	+28.7%
128	55,135 rps · 798% CPU · p99 3.34ms	39,388 rps · 778% CPU · p99 5.68ms	+40.0%
256	56,242 rps · 797% CPU · p99 6.67ms	39,644 rps · 778% CPU · p99 12.83ms	+41.9%

Verdict: the gate passes. Thread-per-core wins every cell on Graviton, +28-42% rps with p99 roughly halved at load. The gain is smaller than on the x86 VM (+54-88%), exactly as the amplified-context-switch hypothesis predicted, but nowhere near reverting — and the 8-worker leg shows the advantage holds as the worker count grows. The Linux platform default stays as shipped in #891.