Mastering Files, File Descriptors, and I/O Multiplexing in Linux

Before diving into I/O multiplexing, we review the notion of a file and a file descriptor in Linux.

What is a file?

In Linux a file is simply a sequence of N bytes (b1, b2, …, bN). All I/O devices—disk, network sockets, terminals, pipes—are abstracted as files, so the same read/write interface works for any device.

Common file‑related system calls are:

open – open a file

seek – change read/write offset

read / write – transfer data

close – close the file

File descriptor

A file descriptor is just an integer that the kernel uses to refer to an opened file. The process does not need to know where the file resides on disk or how it is buffered; the kernel handles those details.

int fd = open(file_name); // obtain file descriptor
read(fd, buff);

When a network connection is accepted, accept() returns a descriptor that can be used to read from and write to the client.

// accept a client connection
int conn_fd = accept(...);
if (read(conn_fd, request_buff) > 0) {
    do_something(request_buff);
}

Why simple blocking I/O does not scale

If a process blocks on a read from one descriptor, it cannot handle other ready descriptors, which is unacceptable for servers that must serve thousands of clients simultaneously.

Creating a thread per client avoids blocking but incurs huge overhead when the number of connections grows.

I/O multiplexing concept

Instead of actively polling each descriptor, the process hands a set of descriptors to the kernel and asks to be notified when any become readable or writable. This “passive” approach lets the kernel monitor many descriptors efficiently.

Linux I/O multiplexing mechanisms

Linux provides three system calls for this purpose:

select

poll

epoll

select

Copies the descriptor set into kernel space (limited to 1024 descriptors) and returns when at least one descriptor is ready, but the application must scan the set to find which one.

poll

Similar to select but removes the 1024‑descriptor limit; still suffers from linear scanning overhead.

epoll

Uses an event‑driven model: the process registers interest with epoll_ctl, the kernel tracks only changed descriptors in shared memory, and when an event occurs the kernel wakes the process with the ready descriptor list, eliminating the need for full scans and copying.

Because of its efficiency, epoll has become the de‑facto standard for high‑concurrency servers on Linux.