Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Performance (NFR-07 / ADR-011)

Performance is a first-class pillar, not a Later learning-only note. That does not authorize fake benches.

Rules

  1. A latency, cycle, or “faster than” sentence is a claim. It needs a probe in the honesty ledger.
  2. Optimize only after a probe shows a cost. Do not rewrite the scheduler or heap “for speed” on a hunch.
  3. QEMU virt numbers are one environment. They are not a Raspberry Pi probe and not a published SPEC run.

Baseline probe (this tree)

CNTPCT_EL0 is already used to timeout the M5 timer observe window. The pillars ratchet (when landed) measures a fixed trivial loop:

  • Serial marker perf: cntpct delta=<n> (fail closed on perf: probe missed).
  • #[test_case] asserts the counter advanced.

That probe proves the physical counter is readable and moves. It does not claim a microsecond budget, interrupt latency, or a comparison to other kernels.

IRQ-to-handler probe (this tree)

When CNTP fires, the handler records CNTPCT − CNTP_CVAL before rearm. After several ticks the hello kernel prints:

  • Serial marker perf: irq-delta min=<a> max=<b> spread=<b-a> n=<n> (fail closed on perf: irq-delta missed).
  • #[test_case] asserts samples exist and max >= min.

That is a spread of IRQ-to-handler counter deltas on this QEMU virt guest. It is not a latency budget, not “faster than X,” and not a published bench. QEMU TCG jitter is one environment.

Host debug ELF size (NFR-08, this tree)

scripts/qemu-smoke.sh prints the host byte size of target/aarch64-ctos/debug/ctos after cargo build:

  • Host marker perf: elf-size bytes=<n> (fail closed if missing or < 4096).
  • This is a measurement, not a size budget and not a “smaller is better” claim.

It does not time QEMU boot.

Boot-to-ready CNTPCT (NFR-08, this tree)

kernel_main samples CNTPCT_EL0 after paging::init (MMU + D-cache on) and again after Hello World! (init complete):

  • Serial marker perf: boot-delta ticks=<n> (fail closed on perf: boot-delta missed).
  • #[test_case] asserts a sample exists and the counter advanced.

That is kernel_main-entry to after-init on this QEMU virt guest. It is not a latency budget, not QEMU startup time, not a published bench, and not criterion.

OS/app slot disconnect (A9) — expected shape, not a bench

A9 #48 would load a separate app payload after the OS image (immutability.md, site measure.md). That is Planned after Track A ABI/loader. Today is still one linked ELF. There is no Verified delta. Do not invent a “faster/slower than linked-in” number.

ClassWhat we expect (hypothesis, unmeasured)Honesty
CostsExtra boot/load work; each SVC crossing; ASID/TTBR0 switches into the app map; optional COW later if payloads are sharedCost only after a probe. Not a budget.
Neutral / winsSteady EL0 compute (once mapped) should look like today’s standing stub, not like a new ISA. Smaller OS updates are an operational win (rebuild kernel without apps), not a CNTPCT winOperational ≠ measured latency.
GateKeep perf: boot-delta. Add a new app-load CNTPCT probe (perf: app-load or similar) when the loader exists. Fail closed on missMarker does not exist today.

Until that app-load probe prints on a two-artifact boot, A9 performance stays Planned. QEMU TCG jitter is still one lab. No criterion crate. No “slot disconnect is free.”

Later probes (Planned)

  • A tighter “first instruction of _start” sample if someone maps a .data slot that BSS-clear will not wipe.
  • App-load CNTPCT (A9) — only after a real loader; see the table above. Site KPI page: measure.md. Do not invent a percent.

Do not add a host criterion crate or a “bench.yml” that prints invented numbers.