Same limits
Each service runs in its own container limited to 2 CPUs and 1 GiB of memory, like a small production pod.
Open source · 15 implementations · 8 languages · 2 published runs
Benchmania measures how software performs under realistic, controlled conditions. Anyone can contribute implementations, benchmark domains and measurements from their own hardware; every number is published with the hardware, versions and raw data behind it.
Featured domain
HTTP services backed by PostgreSQL, each running in the same small, resource-limited container.
| Rank by Throughput index | Implementation | ||||
|---|---|---|---|---|---|
| 1 | TypeScript · BunBunBare | 99.2 | 1.27ms | 75.9MiB | 200ms |
| 2 | Rust · axumTokioMicro | 91.6 | 0.93ms | 21.6MiB | 173ms |
| 3 | C# · ASP.NET Core 10 LTSJIT.NET 10Micro | 73.1 | 1.34ms | 144.1MiB | 501ms |
| 4 | C# · ASP.NET Core 10 LTSNative AOT.NET 10Micro | 65.4 | 1.50ms | 93.2MiB | 690ms |
| 5 | C# · ASP.NET Core 11JIT.NET 11Micro | 59.2 | 1.70ms | 155.6MiB | 496ms |
| 6 | TypeScript · FastifyNode.jsMicro | 56.8 | 1.90ms | 223.5MiB | 582ms |
| 7 | Go · net/httpGoBare | 55.3 | 1.86ms | 41.9MiB | 170ms |
| 8 | C# · ASP.NET Core 11Native AOT.NET 11Micro | 52.6 | 1.78ms | 97.2MiB | 215ms |
| 9 | Java · Spring BootJVMJVM (HotSpot)Full-stack | 45.6 | 6.95ms | 680.0MiB | 2,368ms |
| 10 | Java · Spring BootJVM + AOT cacheJVM (HotSpot)Full-stack | 45.4 | 2.83ms | 671.4MiB | 1,486ms |
| 11 | TypeScript · ExpressNode.jsMicro | 41.7 | 2.53ms | 220.0MiB | 532ms |
| 12 | Elixir · Plug + BanditBEAM (Erlang/OTP 29)Micro | 32.4 | 4.87ms | 156.6MiB | 1,203ms |
| 13 | PHP · FrankenPHPFrankenPHP (worker mode)Bare | 25.5 | 7.78ms | 134.8MiB | 307ms |
| 14 | Java · Spring BootNative imageGraalVM Native ImageFull-stack | 17.0 | 38.56ms | 244.3MiB | 274ms |
| 15 | Python · FastAPICPythonFramework | 14.1 | 14.19ms | 141.6MiB | 1,274ms |
Throughput · the best valid result in each of the 6 workloads behind the throughput index
Community
Implementations, domains and measurements all arrive as pull requests. Run the suite on your own machine and your results are published next to everyone else's, with the hardware they were measured on.
Run the suite on a Linux machine with Docker and open a pull request that adds the result bundle.
Submit a runA language, framework or variant: a directory with an impl.toml, a Dockerfile and the source.
Add an implementationA new area such as databases or compute. Domains are discovered automatically and get their own pages.
Add a domainNo community runs yet. Add the first. Reference runs are measured by the maintainers; every other merged run is labelled community.
Submit a runMethodology
Numbers are only worth comparing when the conditions are identical. Every implementation gets the same box, the same work and the same statistics.
Each service runs in its own container limited to 2 CPUs and 1 GiB of memory, like a small production pod.
Service, database and load generator are pinned to separate cores, so they never compete for CPU time.
SMT is on. The service gets one logical CPU per physical core; the SMT siblings of those cores stay idle.
Several passes in rotating order. Results are medians with a 95 % bootstrap confidence interval; outliers are flagged, never silently removed.
Every implementation must pass functional checks before it is measured, and the backend harness counts the database statements each request really runs.
Go deeper
The leaderboard shows four numbers. The dataset has many more, and all of it is open.
15 implementations, 11 scenarios and 39 metrics with confidence intervals.
Open explorerPut up to four implementations side by side and see the relative difference in every scenario.
Start comparingEvery run records the hardware fingerprint, kernel, Docker version and CPU layout it was measured on.
See machinesDownload the dataset as JSON under CC-BY-4.0, or the raw per-pass result bundles.
Get the data