# Benchmarks

Open **Performance**, pick one or more requests, choose a load profile and press **Start**. The results stream in while the run is going.

![Benchmark results with a timeline and one outcome strip per request](benchmark.png)

## Load profiles

| Profile | What it does |
| --- | --- |
| Fixed | A number of parallel senders share a total number of requests. |
| Ramp | The number of senders grows from a start value to an end value over a duration. |
| Rate | Aims for a number of requests per second, however long each one takes. |
| Watch | Samples on an interval to find outages. See [Watching for an outage](./watch.md). |

## Reading the results

- **Tiles:** requests, min, average, p95, max and failed.
- **Over time:** average, min and max per time slice, with an outcome strip for each request.
- **Distribution:** how the durations spread.
- **Per request:** the same numbers for each request on its own.
- **Previous runs:** earlier runs of the same benchmark, to compare.

Durations below 2 ms are kept to a tenth of a millisecond, so a fast local API still has a distribution. The aggregation keeps no per-response list, so a run of a million responses costs the same to display as a run of a thousand.

## What is sent

A benchmark builds every request exactly as Send does, with the same variables, auth and checks. A failed check counts as a failed request.