#Benchmarking

This page describes the public benchmark surface for IndexBus and how to compare results without overstating what they mean.

IndexBus owns benchmarkable transport, routing, and typed-message patterns in this repository. The primary checked-in harness is the indexbus-bench crate, which is run with Criterion.

#What this page is for

• identify which checked-in benchmark surfaces belong to IndexBus
• connect pattern-level docs to runnable harnesses
• define the minimum context that should travel with any published result

#Benchmark surfaces in this repo

The repo currently exposes three practical benchmark surfaces:

1. pattern benchmarks through cargo bench -p indexbus-bench --bench patterns
2. structured messaging benches through cargo bench -p indexbus-bench --bench structured_messaging
3. stream-oriented benches through cargo bench -p indexbus-bench --bench streams

Those benches live under crates/indexbus-bench and cover typed roundtrips, fanout, transport codec costs, and deployable routing patterns.

#Read these as topology evidence

IndexBus results are sensitive to topology.

• SPSC, MPSC, and fanout shapes should not be compared as if they were the same workload.
• In-process and cross-process paths should not be merged into one headline number.
• Wake strategy, batching, subscriber lag posture, and overrun behavior can dominate the result.

When sharing results, state the topology first and the number second.

#Minimum reporting set

When you publish or compare IndexBus numbers, include at least:

• the bench suite and bench name
• payload shape or codec family
• process boundary, transport choice, and wait strategy
• CPU pinning, NUMA locality, and whether perf counters were captured
• throughput plus latency percentiles

If you benchmark one of the standard lane patterns, use the same pattern names described in Lanes (v1): core execution patterns.

#Interpreting the result

Use benchmark output to answer one of these questions:

• what is the steady-state cost of a specific transport or codec path?
• how does a routing pattern behave under the same host posture?
• where does overrun or wake behavior start to dominate the datapath?

Do not treat a single Criterion run as a support contract. The normative contract remains in the v1 spec, and the operational contract remains in the triage and counter guides.

#Related docs

• Performance & tuning (v1)
• Lanes (v1): core execution patterns
• Ops triage (v1)
• Router counters reference (v1)