Skip to content

This is general knowledge or an external-engine reference, not an Apollo architecture or implementation claim. See Apollo's architecture status for target terminology and verified implementation boundaries.

Q77: What is the difference between IOCP and Epoll? ​

Core Conclusion ​

The biggest difference between IOCP and epoll is not "which is faster", but that the I/O models and programming styles they represent are different:

  • epoll leans more toward readiness notification
  • IOCP leans more toward completion notification

So when comparing them, the focus should not be abstract benchmarks, but rather:

  • Operating system platform
  • Network model
  • Event loop design
  • How the business layer takes over I/O results

One: The Most Core Difference Between the Two ​

1. epoll ​

The typical semantics are:

  • The kernel tells you that some fd is now readable or writable
  • You then go and complete the actual read/write yourself

This means the application layer usually has to handle on its own:

  • Non-blocking reads/writes
  • Partial reads/writes
  • Buffer advancement

2. IOCP ​

The typical semantics are:

  • You first submit an asynchronous I/O request
  • After the I/O completes, you receive a completion event

This is closer to a "completion notification" model.

Two: Why You Cannot Simply Say Which Is More Advanced ​

They serve different system environments:

  • epoll is one of the core mechanisms in the Linux ecosystem
  • IOCP is a mature high-performance asynchronous I/O model under Windows

The real engineering judgment is usually:

  • Use epoll well on Linux
  • Use IOCP well on Windows

Rather than forcing an abstract win/lose comparison.

Three: They Affect Program Structure Differently ​

1. epoll Style ​

More commonly seen in:

  • Reactor pattern
  • Event-loop-driven
  • Non-blocking socket state machines

2. IOCP Style ​

More commonly seen in:

  • Proactor style
  • Submit a request, then wait for the completion event

This directly affects how connection objects, buffers, and state machines are organized.

Four: Real Performance Differences Are Usually Not the Primary Concern ​

In real online systems, what usually matters more is:

  • Whether business logic is too heavy
  • Whether the protocol layer is efficient
  • Whether the thread model is reasonable
  • Whether there is lock contention and invalid copying

In other words, many bottlenecks sit above both, rather than in epoll or IOCP themselves.

Five: How Cross-Platform Systems Should Handle It ​

A more common approach is:

  • Abstract a unified event or completion interface
  • Implement separately per platform at the bottom layer
  • The business layer does not directly depend on platform details

But this abstraction must be moderate; you must not flatten out the core advantages of both models for the sake of "complete unification".

Six: Common Misconceptions ​

1. IOCP Is Definitely Faster Than epoll ​

There is no universally true conclusion; it depends more on the specific implementation and business model.

2. epoll Is Asynchronous I/O ​

More accurately, it is mainly an efficient I/O readiness notification mechanism.

3. Doing Cross-Platform Requires Forcing Unification Into the Same Underlying Semantics ​

Over-unification often sacrifices platform characteristics and implementation clarity.

References ​

  • Materials on epoll, IOCP, Reactor, and Proactor

Released under the MIT License