# Load-generator reference run, 13 August 2026

## Question

How do Apache JMeter, k6, Gatling and Locust differ in generator resource use when they execute the same simple closed workload?

## Workload

- 1,000 virtual users for one hour.
- Closed workload with fixed think time.
- Local Node.js target with a fixed 50 ms response delay.
- Theoretical throughput: 952 requests per second.
- Requests counted independently on the target side.

## Rig

- One 4-core Intel i7-7700 host with 31 GB RAM.
- Target pinned to logical CPU cores 0 and 1.
- Generators pinned to logical CPU cores 2 to 7.

## Engines

- Apache JMeter 5.6.3, heap set to 2 to 4 GB.
- k6 v2.2.0.
- Gatling 3.15.1. The closed workload used an explicit maximum duration.
- Locust 2.46.3 using FastHttpUser and four workers, following Locust's multi-core guidance.

## Results

See `load-engine-benchmark-2026.csv`.

## Limits

This is one run per engine, on one host, with one simple closed scenario and no network distance. There are no repeated-run confidence intervals. The figures compare resource profiles under this workload. They do not establish maximum scale, script maintainability, protocol breadth, cloud economics or production-fleet cost.

At fixed concurrency and think time, throughput is set by the scenario. All four engines delivered 950 to 952 target-side RPS, within a fraction of one per cent of the theoretical 952 RPS. CPU and memory were the meaningful differences in this run.