Feature request: Reduce memory usage of sing-box to make it run on systems with 128MB memory.

[arinc9]

Hello.

We use sing-box on embedded hardware. It runs well on systems with 256MB memory or more. A 128MB system runs out of memory with sing-box under load. We'd be happy to financially support an effort to reduce the memory usage of sing-box under load.

[nekohasekai]

Your description is too vague. Please describe your use case in detail, such as the operating system, the protocol used, and the specific load.

[arinc9]

Hey thanks for your response. We use the latest stable OpenWrt release. We use a single outbound, vmess with MPTCP. The inbound is tproxy so the client's traffic is transparently proxied to sing-box. We let the CPU handle as much traffic as it can.

[megapro17]

what kind of rulesets are you using? my router is using a more than a million domain list and 256 mb ram is almost full

[arinc9]

what kind of rulesets are you using? my router is using a more than a million domain list and 256 mb ram is almost full

We don't use rulesets; we simply proxy user traffic.

[nekohasekai]

L4 proxies are inherently not designed for proxying all traffic. Users from China typically use rules to route local requests to direct, and features like auto_redirect or the new bypass action allow connections to be excluded from sing-box's routing entirely.

If your use case is simply proxying all traffic without any routing rules, WireGuard-based encryption would be a better choice for your scenario.

The current performance characteristics are partly due to Go language limitations. sing-box 1.14 may bring improvements in this area, but that remains a longer-term goal.

[arinc9]

We don't do encryption. And proxying TCP and UDP traffic solves a lot of problems that tunnelling all traffic such as with WireGuard would cause by design. I'd love to see improvements in memory usage however long they may take.

[megapro17]

Hmm, maybe something like this https://github.com/cfal/shoes

[megapro17]

The current performance characteristics are partly due to Go language limitations

Protocol Sniff low performance is also caused by this?

[github-actions]

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this will be closed in 5 days