Here is a patch to optimize ipv4 netmask parsing by maintaining a cache (there are not many valid ipv4 netmask representations). This should be especially useful with #16531.

Updated patch, also optimizing v6 netmask parsing (same principle).

Before patch:

$ ./python -m timeit -s "import ipaddress" "net = ipaddress.IPv6Network(('2001:db8::', 96))"
10000 loops, best of 3: 26.1 usec per loop

$ ./python -m timeit -s "import ipaddress" "net = ipaddress.IPv4Network(('10.0.0.0', 23))"
100000 loops, best of 3: 17 usec per loop

After patch:

$ ./python -m timeit -s "import ipaddress" "net = ipaddress.IPv6Network(('2001:db8::', 96))"
100000 loops, best of 3: 13.8 usec per loop

$ ./python -m timeit -s "import ipaddress" "net = ipaddress.IPv4Network(('10.0.0.0', 23))"
100000 loops, best of 3: 14.3 usec per loop

Why not just use functools.lru_cache?

Because that would incur the cost of LRU logic and locking, which we don't need here.

With C implementation (issue14373) functools.lru_cache is so fast as manually written specialized code.

What I want to say, the patch LGTM, but after committing issue14373 we should simplify the code by using functools.lru_cache().

Actually, using lru_cache(maxsize=None) would enable a simple infinite cache like in the patch. But it's not like a lot of code would be saved.

New changeset 2158614e1607 by Antoine Pitrou in branch 'default':
Issue #21486: Optimize parsing of netmasks in ipaddress.IPv4Network and ipaddress.IPv6Network.
http://hg.python.org/cpython/rev/2158614e1607