[Discussion] Re-evaluating the adoption of io_uring in Netpoll and its positioning vs. Monoio

[NineSue]

Hi developers,

I've been following the performance optimizations in Netpoll for a while. As Linux kernel versions (6.x+) and io_uring become more mature, I have a few small questions regarding the roadmap:

1. Performance Threshold for io_uring: In your previous internal evaluations, what were the specific QPS or latency bottlenecks where epoll fell short of io_uring? With the latest kernels, is there a plan to introduce a formal io_uring branch?

2. Go 1.21+ runtime.Pinner Integration: One of the historical challenges for io_uring in Go was memory pinning/safety. Since Go 1.21 introduced runtime.Pinner, does the team have plans to leverage this feature to resolve the buffer ownership issues in a potential io_uring implementation?

3. Selection Logic (Netpoll vs. Monoio): Given that Monoio (Rust) is built heavily around io_uring, in which specific business scenarios would ByteDance recommend migrating from Netpoll (Go) to Monoio/Volo (Rust)? Is Netpoll's goal to remain as an optimized epoll-based refactor of Go's netpoller, or will it eventually embrace Proactor models?

Looking forward to your insights!

[huzhao37]

这是一封自动回复邮件。已经收到您的来信，我会尽快回复。

[NineSue]

I’ve been diving into Netpoll's architecture and the possibility of io_uring for the past two days. To be honest, I think I’ve found some of my own answers regarding the trade-offs, and I'd like to sync with the "pros" here to see if my direction is correct.

1. The Pinner Performance Trade-off: After reflecting on the runtime.Pinner introduced in Go 1.21, I have a strong feeling that relying heavily on it for a Proactor model might be counter-productive. The overhead of constant pinning/unpinning, along with the added pressure on the GC, could easily negate the performance gains from io_uring. To me, it feels like it somewhat contradicts the "GC-managed" philosophy. How is the team planning to balance memory safety with runtime overhead? Are we considering a Pinned Buffer Pool or some other middle ground?

2. Roadmap & Contribution: The README mentions that io_uring is on the roadmap. I’m genuinely curious about the current design status and the expected timeline. Is there an active experimental branch or an RFC (Design Doc) I could look into? I am very interested in getting my hands dirty and contributing to this part of the project. Could you share the team's current thinking on the starting point for io_uring integration? I'd love to jump in.

Thanks.