Performance Benchmarks

This page documents the benchmark entry points that are present in the source-available repository. Results vary by hardware, operating system, runtime configuration, and feature flags. Treat local results as capacity-planning input, not as a guarantee.

Benchmark Scope

Area	Benchmark
Gateway WebSocket connections	crates/nexis-gateway/benches/websocket_connections.rs
Gateway message throughput	crates/nexis-gateway/benches/message_throughput.rs
Gateway routing	crates/nexis-gateway/benches/routing.rs
Context summarization	crates/nexis-context/benches/context_summarization.rs
Meeting SFU domain	crates/nexis-meeting/benches/sfu_benchmark.rs
Document CRDT domain	crates/nexis-doc/benches/crdt_benchmark.rs

Run Benchmarks

cargo bench --workspace

Run a focused benchmark:

cargo bench -p nexis-gateway --bench websocket_connections
cargo bench -p nexis-gateway --bench message_throughput
cargo bench -p nexis-gateway --bench routing
cargo bench -p nexis-context --bench context_summarization
cargo bench -p nexis-meeting --bench sfu_benchmark
cargo bench -p nexis-doc --bench crdt_benchmark

Reporting Results

When publishing benchmark numbers, include:

• Commit SHA
• Host CPU, memory, operating system, and kernel version
• Rust version
• Build profile and feature flags
• Benchmark command
• Load shape, connection count, message size, and duration
• Whether the run used localhost, container networking, or a remote network

Capacity Planning Notes

• Gateway connection capacity depends on file descriptor limits, memory, CPU, kernel networking settings, and workload shape.
• AI summarization latency depends heavily on the configured provider and model.
• Container and Kubernetes deployments should set explicit CPU, memory, and file descriptor limits before load testing.
• Production claims should be based on repeatable benchmarks in the target environment.

Monitoring During Tests

The gateway exposes Prometheus metrics at /metrics when running:

curl http://localhost:8080/metrics

Use these metrics together with host-level CPU, memory, network, and file descriptor telemetry.