benchmarks

benchmarks

Measures what an operation costs. Not published.

This benchmark uses ProcessTransport, which starts a tmux process for each dispatch. It measures how many dispatches each operation takes and regenerates docs/benchmarks/operation-costs.md from a real run.

$ ./gradlew operationBenchmark -PlibtmuxTmux=/path/to/tmux

The table is never hand-edited, and it stamps which tmux answered, because a table without its conditions is a claim rather than a measurement. Read the dispatch counts; the milliseconds are one machine at one moment.

Four things are measured, and each exists because the library made a choice that costs something: collapsing round trips with batch() and chain(), capturing the hierarchy, fencing a handle's command against a replaced server, and reading a scope's options.

Its own module, and excluded from check: a benchmark starts a tmux server per case and takes seconds. Keeping it inside a published artifact's tests made that a matter of remembering a tag rather than a matter of where the code lives.

The Scala facades' modes

ModeBenchmarks runs equivalent live-tmux workloads through each way the Scala facades reach tmux: Java process reads, direct-style blocking reads, Cats serial and bounded reads, batches and chains, admitted control requests, polling and pushed output. It writes raw JSON samples — elapsed time, allocation, client, process and worker counts, p50/p95/p99 and throughput — wherever it is told:

$ ./gradlew :benchmarks:scalaModeBenchmark \
    -PlibtmuxTmux=/path/to/tmux \
    -PscalaBenchSocket=/tmp/libtmux-java-dev/bench/s \
    -PscalaBenchConfig=/tmp/libtmux-java-dev/bench/empty.conf \
    -PscalaBenchOut=/tmp/libtmux-java-dev/bench/modes.json

It refuses an existing socket and validates pane ids, capture order and completion markers before recording anything. A run describes the machine and revision it ran on, so none is committed.

Next

• Batching and chaining · the measured table