It is too soon to conclude that POE is the cause.
The difference in CPU time between your Python and Perl tests is
0.0002 seconds per article (user+sys). It's not a significant
difference at this scale. It could be caused by any number of low-
level variations between POE and Twisted, or Perl and Python.
The wall times are a better place to look for inefficiencies. They
are three orders of magnitude larger than the total CPU times.
The large difference between wall and CPU times indicates a non-CPU
bottleneck. In your test case, it's probably the network.
What could account for network differences?
You are comparing two different NNTP client libraries. They may
interact differently with the server, which is a likely place to
look. To investigate this, I would scale the tests down to a small
number of articles (2 < count <= 10). I would dump the resulting
network traffic for each test. I would use the dumps to compare
interaction patterns between the two libraries. I would also look at
timings: the times between command & response, the times between one
command and the next, the length of time to receive an article, and so
on. tcpdump and tshark (or wireshark) are very good for this sort of
thing.
If network interaction is comparable, we should revisit the points I
outlined in my last message. They don't seem to have made much of an
impression, but they may still be relevant.
Thank you.
--
Rocco Caputo - [email protected]
On Mar 25, 2009, at 13:18, howard chen wrote:
But anyway, now I re-run the test, instead of downloading 100 posts,
now I download 1K articles.
python
real 3m6.007s
user 0m0.088s
sys 0m0.016s
perl
real 6m13.730s
user 0m0.268s
sys 0m0.024s
