[Boost-users] boost::asio::io_service::poll_once vs EPOLLET

Subject: [Boost-users] boost::asio::io_service::poll_once vs EPOLLET
From: Jason Aubrey (jason_at_[hidden])
Date: 2014-03-11 12:19:44

The conditional define of boost::asio::io_service that uses epoll always seems to use edge triggered (ET) flag instead of level triggered (LT).

This makes me think poll_once should be avoided due to Scenario 1 below.

Can anyone confirm this?

Given:

1) 2 sockets (A and B) added to epoll interface for receiving only

2) 2 UDP messages arrive in the same moment on socket A

3) epoll_wait returns a ready socket (socket A)

Scenario 1: Using ET with poll_once:

4) If io_service::poll_once is called 1 of 2 messages are read

5) epoll_wait(ET) blocks until another message is received

Result: Bad, starvation if another message is not received

Scenario 2: Using ET with poll:

4) If io_service::poll is called 2 of 2 messages are read

5) epoll_wait(ET) blocks until another message is received?

Result: Good, all messages processed without A/B fairness (unfair since read was greedy)

Scenario 3: Using LT with poll_once:

4) If io_service::poll_once is called 1 of 2 messages are read

5) epoll_wait(LT) returns a ready socket (socket A)

6) If io_service::poll_once is called 1 of 1 messages are read

Result: Good, all messages processed with A/B fairness

I don't think Scenario 3 is currently supported by boost::asio::io_service.

Jason