[Windows] speedup connections

[giampaolo]

...a variant of #1676.

[giampaolo]

@mlabonte I am benchmarking this PR with:

import psutil, socket, time

ls = []
for x in range(5000):
    s = socket.socket()
    s.bind(('127.0.0.1', 0))
    s.listen(4)
    ls.append(s)

t = time.time()
for x in range(100):
    psutil.net_connections('tcp')
print(time.time() - t)
[x.close() for x in ls]

The results I get are worse than #1676. Without patch I get 3.35 secs, with patch 3.00, meaning around 9% speedup.

[mlabonte]

I cannot understand, I did the same on my workspace and 1679 is slighly faster than 1676. I even found one case where it's definively faster but still, I have no idea why as both PR do the same operations.

(pr1679) $ python -m timeit "import psutil; psutil.Process().connections(kind='tcp')"
1000 loops, best of 3: 439 usec per loop

(pr1676) $ python -m timeit "import psutil; psutil.Process().connections(kind='tcp')"
1000 loops, best of 3: 636 usec per loop

Maybe the second call is faster than the first in the case of IPv6.

[giampaolo]

Can you get similar results as #1676 (comment)? This one:

Worsth case (end of a quick rampup from 0 K to 15 K connections):
$ python -m timeit -n 1 -r 1 "import psutil; psutil.Process().connections(kind='tcp4')"
1 loops, best of 1: 25.5 sec per loop <--- psutil==5.6.7
1 loops, best of 1: 259 msec per loop <--- pull request

...looks too big of a speedup. Are you sure you were measuring the right thing?

[mlabonte]

Yes, I get similar results comparing master to 1679. It's a particular case: from approx. 9 K conxs, if the rate of new connections is big enough, the size of the table was always bigger than the previous call so we were stuck in the loop. I could get a single call blocked for minutes with ramping up to 60 K connections.

[giampaolo]

Whatever measure I try I always get around 10% speedup. Anyway, it's fine. Merging.

[mlabonte]

Thanks for merging this! :-) Do you have an ETA to release it in pypi?