The harness had a single endpoint returning a 12-byte set_content() body.
That is the one case where the response line, the headers and the body
already share a single write(), so any change to the write path measured
as noise. Comparing a gather-write branch against its merge base reported
0.993x at p = 1.000 while the same branch moved static-file throughput by
a quarter and TLS throughput by nearly half in both directions.
The server now also serves a large set_content() body and small and large
files from a mount point, over HTTPS when a certificate is given, with
--path, --large-mib and --tls selecting the combination.
ab.sh now compiles the harness from the invoking worktree instead of each
ref's own copy, so both refs run an identical workload and a ref that
predates a harness change stays measurable. Only httplib.h varies, through
-I. --timeout is exposed because bombardier's 2s default aborts large TLS
responses, which then fails the non-2xx check.
The existing test_benchmark asserts a single request completes within
5ms, which catches gross connection-setup regressions but cannot see
throughput changes: the effects we care about are tens of microseconds
per request, a hundredth of that resolution. There was no way to answer
"does this patch make the server faster" other than measuring by hand.
Absolute req/s is not usable for that. Running the same binary five
times on an idle 8-core machine gave 53.9k to 74.6k req/s, and shared CI
runners are noisier still, so a number printed per push says nothing.
Build both refs and measure them alternately in one session, flipping
the order each round to cancel ordering bias, then report only the ratio
of the medians. Whether that ratio means anything is decided by an exact
permutation test rather than by comparing the change against the min/max
spread - a single slow round is enough to make a spread check give up,
while the rank test rides it out.
Validated against a patch that removes a redundant poll() per request:
individual measurements ranged 41.7k-93.9k req/s, yet nine rounds
resolved a 1.244x speedup at p = 0.019. The same data truncated to five
rounds was inconclusive, so the workflow defaults to nine.
Manual dispatch only, Linux and non-SSL for now, and it never fails the
build - this is a measurement, not a test.
