CLN/PERF: remove used functions; use C skip list for rolling median

[kawochen]

removes some unused code
reverts this commit a40226e

performance consideration

import pandas
import numpy
arr = numpy.random.rand(1000000)
%timeit pandas.rolling_median(arr, 1000)

master

1 loops, best of 3: 4.94 s per loop

branch

1 loops, best of 3: 821 ms per loop

[jreback]

add the pr number (as the issue number)

[jreback]

are sure about this?

[kawochen]

I think it makes clean-up code cleaner. I will write something to test for leaks a bit later.

[jreback]

you may be able to insert nogil (at the top-level), and make almost all of skiplist.pyx nogil (obviously where you raise you have re-acquire).

[kawochen]

In [2]: import pandas
import numpy
arr = numpy.random.rand(1000000)

In [4]: from pandas.util import testing

In [5]: @testing.test_parallel(2)
def g():
    pandas.rolling_median(arr, 1000)
   ...:     

In [6]: @testing.test_parallel(1)
def f():
    pandas.rolling_median(arr, 1000)


In [7]: %timeit f()
1 loops, best of 3: 752 ms per loop


In [8]: %timeit g()
1 loops, best of 3: 972 ms per loop

[jreback]

can you add an asv benchmark? not sure we have much for rolling in general......

[jreback]

say Improved performance of ....

[kawochen]

OK. might as well release the GIL on roll_*

[jreback]

yep - same or another

[kawochen]

Added asv benchmarks for the gil release. But I can't get any of those to show in asv bench at all. Have I written the tests incorrectly?
I do asv continuous upstream/master HEAD -b gil while on my branch.

A simple timeit does show the improvement, e.g.
for rolling_mean

import pandas
import numpy
from pandas.util.testing import test_parallel
arr = numpy.random.rand(1000000)
@test_parallel(num_threads=2)
def f():
     pandas.rolling_mean(arr, 100)
%timeit f()

branch

100 loops, best of 3: 12.5 ms per loop

master

10 loops, best of 3: 19.5 ms per loop

for rolling_kurt
branch

10 loops, best of 3: 65.5 ms per loop

master

10 loops, best of 3: 98.8 ms per loop

the time_rolling_median in stat_ops.py shows the same 5x improvement I get using timeit.

[jreback]

thanks!