It is a bit tricky to benchmark WebSocket libraries, since you're going to be limited by the speed of the slower end (i.e. benchmarking a server requires an even faster client and vice-versa). The uWebSockets library for C/C++ comes with a benchmarking tool called load_test that should be fast enough not to bottleneck all currently known WebSocket library implementations and has been used for these benchmarks.
I run these benchmarks irregularly on my main system (CPU: AMD Ryzen 9 7950X, RAM: 32GB). The kernel version last used is:
$ uname -a
Linux ceres 6.7.0-0.rc3.20231129gt18d46e76.329.vanilla.fc39.x86_64 #1 SMP PREEMPT_DYNAMIC Wed Nov 29 06:20:06 UTC 2023 x86_64 GNU/Linux
and the benchmarks are run in a podman container based on Alpine Linux Edge.
We're comparing some popular Rust WebSocket libraries to tokio-websockets and uWebSockets as a reference. All of the implementations are asynchronous. For benchmarking we use a simple plaintext echo server.
In order to bottleneck faster, we're also using a single-threaded runtime.
The following libraries are compared (in alphabetical order):
- fastwebsockets (unsound and not thread-safe, non-strict spec compliance)
- rust-websocket (uses EOL tokio 0.1, non-strict spec compliance)
- tokio-tungstenite (non-strict spec compliance)
- tokio-websockets (strict spec compliance)
- uWebSockets (non-strict spec compliance)
The benchmarking script was taken from fastwebsockets and modified.
