# Proxy Mode Performance
This page records local end-to-end proxy mode benchmark results from `mode_perf`.
The benchmark starts a local HTTP target, proxy endpoints, and a client path,
then measures each proxy mode. Non-WG modes use the SOCKS request path. WG
uses separate target/server/client child processes plus direct tunnel-IP HTTP
requests, and is disabled by default because it creates real TUN/device
interfaces.
Run it with:
```bash
cargo bench --bench mode_perf
```
To run the full suite, including the WG device/noise/obfs profiles that require
host privileges:
```bash
sudo make perf
```
Default workload:
- Warmup: 100 requests
- Small request run: 1000 new TCP/SOCKS requests
- Large response run: 8 downloads of 1 MiB each
- Host: localhost loopback
- Build profile: Cargo bench/release profile
## Latest Result
Measured on 2026-04-17.
Command:
```bash
sudo make perf
```
| native-http | 2064.6 | 0.48 | 0.48 | 0.51 | 631.2 | 1.00 MiB | |
| native-mux | 3365.1 | 0.30 | 0.26 | 0.38 | 540.2 | 1.00 MiB | |
| daze-ashe | 3148.2 | 0.32 | 0.29 | 0.40 | 294.2 | 1.00 MiB | |
| daze-baboon | 2842.1 | 0.35 | 0.31 | 0.44 | 306.2 | 1.00 MiB | |
| daze-czar | 3425.7 | 0.29 | 0.27 | 0.34 | 206.8 | 1.00 MiB | |
| wg | 2545.7 | 0.39 | 0.39 | 0.44 | 43.8 | 1.00 MiB | real TUN/device |
## WG Mode
WG mode is opt-in when running the bench binary directly. The benchmark driver
spawns separate child processes for the HTTP target, WG server, and WG client,
assigns tunnel IPs, and measures HTTP requests to the server tunnel IP from the
parent process. The Makefile target includes every WG profile by default:
```bash
sudo make perf
```
To run only WG:
```bash
sudo make perf RUNNEL_PERF_MODES=wg
```
To compare only the noise engine without and with obfuscation:
```bash
sudo make perf RUNNEL_PERF_MODES=wg RUNNEL_PERF_WG_PROFILES=noise,mask
```
Optional WG-specific settings:
```bash
sudo make perf \
RUNNEL_PERF_MODES=wg \
RUNNEL_PERF_WG_CLIENT_IP=10.88.0.2 \
RUNNEL_PERF_WG_SERVER_IP=10.88.0.1 \
RUNNEL_PERF_WG_MTU=1420 \
RUNNEL_PERF_WG_READY_TIMEOUT=15
```
Device names can also be overridden:
```bash
sudo make perf \
RUNNEL_PERF_MODES=wg \
RUNNEL_PERF_WG_SERVER_DEVICE=runnelwgs0 \
RUNNEL_PERF_WG_CLIENT_DEVICE=runnelwgc0
```
On macOS the default device setting is `auto`. On Linux the default managed
device names are `runnelwgs0` and `runnelwgc0`.
## Tuning
The benchmark can be adjusted with environment variables:
```bash
RUNNEL_PERF_REQUESTS=500 \
RUNNEL_PERF_LARGE_DOWNLOADS=16 \
RUNNEL_PERF_LARGE_BYTES=4194304 \
cargo bench --bench mode_perf
```
Available variables:
- `RUNNEL_PERF_MODES`: comma-separated mode list, for example `native-http,daze-czar`
- `RUNNEL_PERF_WG`: include WG in the default mode list when set to `1`
- `RUNNEL_PERF_WG_PROFILES`: comma-separated WG profile list, for example `noise,mask`; `all` includes `device`, `noise`, `mask`, and `stealth`
- `RUNNEL_PERF_WARMUP`: warmup request count
- `RUNNEL_PERF_REQUESTS`: small request count
- `RUNNEL_PERF_LARGE_DOWNLOADS`: large download count
- `RUNNEL_PERF_LARGE_BYTES`: bytes per large response
- `RUNNEL_PERF_LOG=1`: enable benchmark logging
- `RUNNEL_PERF_WG_CLIENT_IP`: WG client tunnel IP, default `10.88.0.2`
- `RUNNEL_PERF_WG_SERVER_IP`: WG server tunnel IP, default `10.88.0.1`
- `RUNNEL_PERF_WG_CLIENT_DEVICE`: WG client device name
- `RUNNEL_PERF_WG_SERVER_DEVICE`: WG server device name
- `RUNNEL_PERF_WG_MTU`: WG tunnel MTU, default `1420`
- `RUNNEL_PERF_WG_READY_TIMEOUT`: seconds to wait for the WG target to become reachable
## Notes
These numbers are useful for comparing modes on the same machine and build,
not as absolute network throughput claims. The benchmark is intentionally
localhost-only to reduce external network noise.
WG is omitted from the default end-to-end run because the production WG mode
creates real TUN/devices. The opt-in WG benchmark is intended to be run
manually with `sudo`.
