Rust Websockets

Comparing the default axum , axum-tws and tungstenite (with 1,024 sized buffer).

Adding into the comparions, different memory allocators.

We first measure idle, then 10,000 connections, then we disconnect all connections.

Measurments are in MB done via memory-stats crate with always_use_statm feature.

Default

Example	Idle	10K	Diconnect
axum-example	6.62	1334.125	45.46
axum-tws-example	6.62	78.0	77.86
tungstenite-example	5.37	69.5	49.88

libc malloc_trim(0)

Example	Idle	10K	Diconnect
axum-example	6.5	1334.5	9.98
axum-tws-example	6.37	78.12	9.54
tungstenite-example	5.37	51.91	16.67

mimalloc mi_collect(true)

Example	Idle	10K	Diconnect
axum-example	7.0	1294.12	416.96
axum-tws-example	6.87	72.75	54.66
tungstenite-example	5.87	44.69	36.86

jemalloc background_threads

Example	Idle	10K	Diconnect
axum-example	9.75	1332.25	24.26
axum-tws-example	9.37	73.73	15.55
tungstenite-example	7.87	47.24	14.19

Summary

• Using tungstenite directly isn't a real option with axum due to the API it requires.
• The default axum setting uses 128K buffer which is too big for more cases, making axum-tws a better option (tried changing axum-tws config but saw no real diferences in results).
• The difference between axum and axum-tws is over 1.3GB versus 77MB , very significant.
• malloc_trim(0) seems to be able to reclaim the most memory but jemalloc is known to help in memory fragmentation and not far behind in performance, while mimalloc is lagging considerably.