rustle against 3proxy
Same machines, same rules, one core each. rustle is first in 41 of 41 tests of the campaign; 16 of them run on both proxies, with the conditions next to every number.
2.6×
more requests per second at 1 KB
55,258 against 21,032 req/s
2.6×
less CPU per request
18.1 against 47.6 µs/req
26×
less memory with 25,000 open tunnels
95 against 2,461 MB
1 core per proxy · authentication on · Scaleway Elastic Metal, Paris, AMD EPYC 8224P · 25 Gbit/s private network (23.5 measured) · median of 3 runs · seed 2004113323 · campaign of 2026-10-28.
The short version
rustle had the better value in 16 of the 16 tests where both it and 3proxy ran. 3proxy was not ahead in any of them. In this campaign 3proxy has no HTTPS proxy port and HTTP/2 to clients, so the tests that need them are not in the tables. Every value is a median of 3 runs, with the exact figures in the tables below.
What each one supports
| Protocol | rustle | 3proxy |
|---|---|---|
| HTTP proxy and CONNECT tunnels | Yes | Yes |
| HTTPS proxy port (TLS) | Yes | No |
| HTTP/2 to clients | Yes | No |
| WebSocket upgrades | Yes | Yes |
As built and configured in the campaign, see “How 3proxy was run” below.
Test by test
Best clean concurrency level of each proxy. The last column is the ratio of rustle’s figure to 3proxy’s: how many times more or less. For memory and CPU, less is better.
Plain HTTP proxy port
| Test | rustle | 3proxy | Ratio |
|---|---|---|---|
| HTTP, 1 KB responsesreq/s | 55,258 | 21,032 | 2.6× more |
| HTTP, 10 KB responsesreq/s | 40,783 | 19,169 | 2.1× more |
| HTTP, 1 KB responses, a new connection per requestreq/s | 27,827 | 5,132 | 5.4× more |
| HTTP, 1 MB responsesreq/s | 2,624 | 1,784 | 1.5× more |
| HTTP, 10 MB responsesreq/s | 268 | 191 | 1.4× more |
| CONNECT tunnels, bulk transferGbit/s | 22.0 | 15.7 | 1.4× more |
| CPU cost per request (1 KB)µs/req | 18.1 | 47.6 | 2.6× less |
| Memory with 1,000 open tunnelsMB | 14 | 104 | 7.4× less |
| Memory with 5,000 open tunnelsMB | 40 | 499 | 12× less |
| Memory with 10,000 open tunnelsMB | 54 | 992 | 18× less |
| Memory with 25,000 open tunnelsMB | 95 | 2,461 | 26× less |
WebSocket, plain
| Test | rustle | 3proxy | Ratio |
|---|---|---|---|
| WebSocket, 64-byte messagesmsg/s | 58,506 | 33,136 | 1.8× more |
| WebSocket round trip, 64-byte messages (p99)ms | 1.3 | 3.2 | 2.5× less |
| CPU cost per WebSocket messageµs/msg | 17.1 | 30.2 | 1.8× less |
| WebSocket, 64 KB messagesGbit/s | 11.6 | 7.9 | 1.5× more |
| WebSocket, one connection per exchangemsg/s | 10,740 | 251 | 43× more |
These are the figures of one campaign, on one kind of traffic. What the numbers do not show, multi-core results and the other proxies
How 3proxy was run
3proxy 1.0.0
from source, Makefile.Linux
- proxy -a -osTCP_NODELAY -ocTCP_NODELAY (documented for small exchanges; -a = no Forwarded header, like the others)
- maxconn, nscache, timeouts
What the campaign showed
- Its best level at 1 KB is 256 connections: it used 19 % of a core at 64 connections and 99 % at 1,024 in the A/B of FAIRNESS.md, hence the concurrency sweep.
- Without any socket option it sits in Nagle and a 40 ms delayed ACK (about 1.5k req/s at 1 KB in the A/B): the default config measures a missing documented option, which is why the options above are set.
- Supports WebSocket upgrades but spends about 4 ms of CPU per new connection.
- No TLS or HTTP/2 listener.
- Thread-per-connection: uses the n cores of the multi-core curve without a core-count setting.
Found a setting that makes 3proxy faster? Open an issue on the proxy repository: the campaign is built to be re-run.
Try rustle
Open source, fast, optimized, safe.