epoll_wait blocks for longer then timeout

Question

My application waits in epoll_wait much longer then I specified in timeout:

22578 09:33:46.959791 epoll_wait(5,  <unfinished ...>
22578 09:33:50.010794 <... epoll_wait resumed> [], 128, 1498) = 0
...
22034 09:35:07.686896 epoll_wait(5,  <unfinished ...>
22034 09:35:09.482526 <... epoll_wait resumed> [{EPOLLIN, {u32=151458248, u64=151458248}}], 128, 362) = 1
...
22036 09:35:41.433241 epoll_wait(5,  <unfinished ...>
22036 09:35:43.176881 <... epoll_wait resumed> [], 128, 97) = 0

I have asked to timeout for one value, but timing shows that it waits for much longer interval.

Time from logs is reliable - I got it from strace output. What can I do to make timeout in epoll more fine grained? And why epoll_wait is SO MUCH inaccurate?

Answer 1

The documentation says:

Note that the timeout interval will be rounded up to the system clock granularity, and kernel scheduling delays mean that the blocking interval may overrun by a small amount.

In other words, the timeout is a minimum value.

But such large delays do not happen when the machine is idle. If you have real-time constraints, ensure that your code can actually run, ie, set the appropriate scheduling policy, prevent it from being swapped out, etc.