Currently, epoll.poll() allocates an epoll_event buffer every time/ this is bad, because it changes an O(number of ready FDs) syscall into a O(maxevents/FD_SETSIZE) complexity.
The patch attached allocates a epoll events buffer per epoll instance, and it only gets reallocated when necessary (note that the reallocation heuristic will probably be improved in #16873 (and having a per instance maxevents count will likely be useful).

Here's a benchmark without patch:
$ ./python -m timeit -s 'import select; ep = select.epoll(); ep.register(1, select.EPOLLOUT)' 'ep.poll()'
100000 loops, best of 3: 4.25 usec per loop

With patch:
$ ./python -m timeit -s 'import select; ep = select.epoll(); ep.register(1, select.EPOLLOUT)' 'ep.poll()'
100000 loops, best of 3: 3.38 usec per loop

Updated to reflect Richard's comment.

New changeset be8e6b81284e by Charles-François Natali in branch 'default':
Issue #16876: Optimize epoll.poll() by keeping a per-instance epoll events
http://hg.python.org/cpython/rev/be8e6b81284e

New changeset 30eb98c8afef by Charles-François Natali in branch 'default':
Issue #16876: Revert be8e6b81284e, which wasn't thread-safe: wait until a
http://hg.python.org/cpython/rev/30eb98c8afef