DataFrame constructor speed

[dieterv77]

Hi, i was playing around with constructing DataFrame's from nested dicts, and noticed that things have gotten a bit slower since v0.6.1.

Here's some sample code i was playing with:

import time
import pandas

print pandas.version

data = dict((i,dict((j,float(j)) for j in xrange(100))) for i in xrange(5000))
t0 = time.time(); df = pandas.DataFrame(data); t1 = time.time(); print t1 - t0

With version 0.6.1, the printed time is about 0.21s on my machine, with a little help from git bisect,
i found that:

commit f3ca67d takes it from 0.21s to 0.44s
commit 9d65e8e takes it from 0.44s to 0.54s

It's possible that some of these were unavoidable considering they may have been necessary bugfixes, but i wanted to
see if anyone else is seeing this too.

environment info: 64bit ubuntu 11.10, python2.7, numpy 1.6.1, cython 0.15.1

[wesm]

Well rats, and you see in those commits I totally thought I was making things faster! I see the issue and I'm going to address it now and add a vbenchmark (http://pandas.sourceforge.net/vbench.html) so we can track the performance more systematically going forward.

[wesm]

OK I fixed things up and even made things a little faster.

before (3ed22d7):

In [3]: timeit df = DataFrame(data)
1 loops, best of 3: 690 ms per loop

after (HEAD):

In [3]: timeit df = DataFrame(data)
10 loops, best of 3: 167 ms per loop

and 0.6.1:

In [3]: timeit df = DataFrame(data)
1 loops, best of 3: 273 ms per loop

Note this problem only affected integer-indexed data. The issue if you're interested had to do with boxing of int64 scalars (from the Index when doing dict lookups).