Core speed¶

How long the core's hot paths take, with no simulator around them (#620). The benchmark results elsewhere measure the protocol in ns-3. Their wall-clock (##PERF##, #131) mixes the core with ns-3's PHY and MAC and moves with every scenario knob, so it cannot show that a core change made AntHocNet slower. These numbers can.

Cost per operation of each core benchmark case, on each merge to main

Reproduce: make reproduce FIG=core-speed re-renders it from the committed data and diffs the PNG (#621).

The cases¶

case	what one operation is
`pheromone_update`	one regular-pheromone deposit on a table with 16 neighbours and 64 destinations, as a backward ant does at each node
`select_next_hop_4` / `_16` / `_64`	one stochastic next-hop choice for a data packet (`betaData`), with 4, 16 or 64 neighbours
`codec_encode_full_ant`	encoding a forward ant carrying the wire maximum of 100 visited hops (1241 bytes)
`codec_decode_full_ant`	decoding that frame
`history_dedup`	recording a new `(src, seq)` in a full dedup tracker (`maxHistory` = 4096), which evicts the oldest entry

Running it¶

make bench-core                       # table: median ns/op, min, max
make bench-core BENCH_ARGS=--csv      # what CI records
make bench-core BENCH_ARGS=--quick    # shorter rounds, for a smoke run

Each case is calibrated to about 0.2 s per round and measured over seven rounds. The median is reported, with the fastest and slowest round beside it.

How CI uses it¶

core-bench.yml runs on every PR and every merge that touches core/:

On a PR, it builds the base branch and the PR on the same runner and runs them interleaved, three times each. It compares the lowest median per case and puts the table in the job summary. A case more than 25 % slower gets a warning annotation. It never fails the PR: shared runners are noisy.
On main, it appends one run to results/core-bench/history.csv with the commit and the runner's CPU model, and redraws the chart above.

Numbers compare only within one CPU model. GitHub-hosted runners change model without notice, and a dashed line on the chart marks each change. The first row was measured in a development container (an Intel Xeon at 2.80 GHz), not on a runner, and is labelled that way in the CSV.