Performance

Performance is one of Ultimo's two headline pillars (alongside security). Our approach is deliberately un-flashy: measure honestly, guard against regressions, and let you reproduce everything. No cherry-picked bar charts.

Why Ultimo is fast

• Built on Hyper 1.0 + Tokio — the same battle-tested async HTTP core that underpins the fastest Rust servers. Ultimo is a thin, well-organized layer over it, not a re-implementation.
• 100% safe Rust, zero unsafe (#![forbid(unsafe_code)]) — native, compiled, no garbage collector and no runtime reflection. Safety and systems-level speed.
• O(1) static routing — route lookup is constant-time regardless of how many routes you register (see below).
• Low per-request overhead — the framework layer (routing, dispatch, response building) is a small, measured constant on top of Hyper.

What we measure

Benchmarks come in two tiers, for two purposes (full detail in BENCHMARKS.md):

1. Framework-overhead micro-benchmarks (criterion, in-process via Ultimo::oneshot). These isolate Ultimo's own cost from network and OS noise, so they're low-variance and reproducible — the basis for our regression gate.
2. End-to-end load benchmarks (oha over real HTTP). Real requests-per-second, for cross-framework comparison — only meaningful on controlled hardware, never a shared CI runner.

We don't conflate the two, and we don't publish cloud-VM throughput as if it were a controlled measurement.

A proven result: O(1) routing

Our benchmark suite earned its keep the day it landed. It revealed that route matching scaled linearly with the number of registered routes — an O(N) scan. We fixed it by indexing static routes in a hash map; lookup is now constant-time:

Found and fixed via the suite itself (#89), with the benchmarks proving the improvement. That's the whole point of measuring.

Regression-guarded

Every pull request that touches the framework runs the micro-benchmark suite against the PR's base commit on the same runner and reports the delta. A change that makes Ultimo slower shows up in review — performance is part of the contract, not an afterthought.

Reproduce it yourself

Framework-overhead micro-benchmarks:

make bench
# or just the HTTP-overhead suite:
cargo bench -p ultimo --bench http_bench

End-to-end throughput on your hardware (the only numbers you should trust for your workload):

# Run a release build of your app, then load it from a separate machine/core set:
oha -z 30s -c 100 --no-tui http://127.0.0.1:3000/

For a fair cross-framework comparison, run the identical endpoint and load profile against each framework on the same hardware in the same session — see BENCHMARKS.md for the full protocol. Published head-to-head numbers will follow once they're generated on dedicated hardware; until then, the recipe above lets you verify on the box that matters to you.