Add make perf and fd table constant lookup test

Draft
CH
6mo ago
+188-7

(This PR is built on top of #68. Please review/merge that one first)

This PR serve as the reference for https://github.com/sysprog21/kbox/issues/38#issuecomment-4415459055


Resolve test gap 3 in #38:

  • Add a check-perf build target that compiles the unit test suite with -O2 -DKBOX_PERF_ONLY in a separate
    binary (test-perf), keeping it independent from the check-unit artifact. Sanitizer and debug flags are stripped from PERF_TEST_CFLAGS so timings reflect optimized production code.

  • Add a benchmark that exercises kbox_fd_table_find_by_host_fd for both present and absent host FDs across table sizes {64, 256, 1024, 4096, 16384} and asserts that per-lookup latency stays within 2x of the N=64 baseline.

Results

--- O(1) Characteristic Perf Test ---
N = 64    | Present:   1.65 ns/op | Absent:   1.39 ns/op
N = 256   | Present:   1.87 ns/op | Absent:   1.38 ns/op
N = 1024  | Present:   1.60 ns/op | Absent:   1.29 ns/op
N = 4096  | Present:   1.63 ns/op | Absent:   1.31 ns/op
N = 16384 | Present:   1.59 ns/op | Absent:   1.39 ns/op
ok

The lookup time is flat across all table sizes — present lookups stay in the 1.59–1.87 ns range and absent lookups stay in the 1.29–1.39 ns range regardless of whether the table has 64 or 16384 entries, adhering to O(1) signature.


Summary by cubic

Adds a perf-only test runner (check-perf) and an O(1) reverse-lookup benchmark for the FD table, plus a fix that correctly downgrades MULTI mappings when aliases close. Covers test gap 3 in #38 and runs in CI without sanitizers to reflect production timing.

  • New Features

    • make check-perf: builds a separate tests/unit/test-perf with -O2 -DKBOX_PERF_ONLY (no sanitizers/debug) so only perf benchmarks run; included in make check and a new CI perf-tests job.
    • O(1) benchmark for kbox_fd_table_find_by_host_fd across {64, 256, 1024, 4096, 16384}; asserts present/absent lookups stay ≤2x the N=64 baseline.
  • Bug Fixes

    • Add host_fd_refs[] and update rev_host_set/clear to maintain counts and downgrade from MULTI to the surviving VFD when one alias remains; counters initialized in kbox_fd_table_init.
    • New regression test (test-fd-table-refcount.c) ensures reverse lookup returns the survivor after other aliases close.