Why epoll Beats select/poll: Inside Linux’s High‑Performance I/O Engine

Overview

Developers using the high‑performance PHP container Workerman know that to support massive concurrent connections they must install the event extension, which is based on libevent – a lightweight, event‑driven networking library.

Libevent is a cross‑platform I/O event notification library. It wraps kernel mechanisms such as Linux’s epoll, Windows IOCP, and BSD kqueue, offering a uniform API for event‑driven programming.

What is epoll?

epoll

is Linux’s scalable I/O event notification mechanism introduced in kernel 2.5.44 to replace the older POSIX select(2) and poll(2) system calls. Unlike those calls, which have O(n) complexity, epoll achieves O(log n) by storing monitored file descriptors in a red‑black tree.

epoll is analogous to FreeBSD’s kqueue; both are built on configurable kernel objects presented to user space as file descriptors. When an event is registered, epoll adds it to its red‑black tree and a callback; ready events are placed on a ready list.

Key Efficiency Points of epoll

Internally manages file descriptors with a balanced red‑black tree, ensuring fast insert, delete, and lookup operations.

When a file descriptor becomes ready, the kernel invokes file_operations->poll to schedule the appropriate callback, enabling event‑driven execution.

The core data structures are the red‑black tree and the ready list; traversing the ready list yields all ready descriptors without scanning the entire set.

Which File Descriptors Can epoll Manage?

Not every file descriptor can be placed into an epoll set because many file systems (e.g., ext2, ext4, XFS) do not implement the poll operation. Only descriptors whose underlying file operations provide a poll method are eligible.

Who Supports It?

The most common eligible descriptor is a network socket. In Linux, sockets implement socket_file_operations (see net/socket.c) which includes a .poll handler.

static const struct file_operations socket_file_ops = {
    .read_iter  = sock_read_iter,
    .write_iter = sock_write_iter,
    .poll       = sock_poll,
    // ...
};

Because sockets provide a poll callback, their file descriptors can be managed by epoll.

Other Supported Descriptors

eventfd : Created via the eventfd system call, this descriptor is used solely for event notification (e.g., producer‑consumer patterns).

timerfd : Created with timerfd_create, it generates readable events when a timer expires.

Summary

Basic I/O multiplexing is a 1‑to‑many model where a single loop handles many file descriptors.

True efficiency requires kernel‑level support; user‑space alone cannot precisely capture I/O readiness.

File descriptors must be set to non‑blocking mode.

epoll’s internal red‑black tree and ready‑list, combined with the kernel’s poll registration mechanism, provide high‑performance fd event management for concurrent I/O.

The full name of epoll is eventpoll; it is implemented as a file‑system‑like kernel module, so saying “epoll is a file system” is not inaccurate.

Socket fd, eventfd, and timerfd all implement the poll interface and can be registered with epoll_ctl for multiplexing.

Traditional file systems such as ext2, ext4, and XFS lack a poll implementation, so they cannot be directly managed by epoll; however, libraries like libaio can be used to bridge file I/O with epoll notifications.