We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·Jun 20·5 min read
Epoll vs. Io_uring in Linux
TL;DR
Linux developers faced inefficiencies with epoll, which required multiple system calls for I/O operations. They built a reverse proxy server called TinyGate, first using epoll and later switching to io_uring for better performance.
✦ Why It Matters
Engineers can leverage io_uring for more efficient asynchronous I/O, reducing overhead in high-performance applications.
Key Takeaways
- Consider using io_uring for high-performance applications.
- Minimize system calls to reduce overhead.
- Explore batch processing capabilities of io_uring.
- Evaluate the trade-offs of using dedicated kernel threads.
Full Summary
Asynchronous I/O management in Linux has traditionally relied on epoll, which notifies developers when I/O is possible but requires multiple system calls for each operation, leading to high overhead. In response to performance limitations, a reverse proxy server named TinyGate was developed, initially using epoll and later rewritten to utilize io_uring, a newer interface introduced in 2019.
Unlike epoll, io_uring informs developers when I/O is completed and allows batch processing of operations, significantly reducing the number of system calls. The implementation of io_uring led to a dramatic performance improvement, although specific benchmark numbers were not provided.
Additionally, io_uring can be configured to minimize system calls further by using a dedicated kernel thread for polling. This evolution in I/O handling techniques highlights the importance of optimizing system calls for high-performance applications.