I have a few issues with the way the benchmarks are run, and then some
actual useful comments. :)
time(1) includes Perl & Python startup/shutdown time, which may be
significant.
Your times seem to be wall clock based, which can vary based on system
load, network conditions, and so on. Show the CPU times instead.
It appears that you've only run one iteration. Benchmarks should be
run enough times to account for statistical variance.
Ok, that said:
A deep and accurate answer requires a deep knowledge of all
technologies involved. I'm not familiar with Python and Twisted, so I
can't make a good comparison by myself.
A previous reply on this thread suggested that CORE::select() was
taking the most time, which strongly implies that POE spends more time
waiting for network activity than Twisted does.
If that's the case, it may be that POE may be checking for I/O faster
than Twisted, and therefore more often, which could result in more and
smaller reads from the socket. The overhead per read will be higher,
although the reads will be more timely.
Someone with dual Perl/Python and POE/Twisted experience would be best
to dig into this to find the actual cause. We can then act on that,
rather than on conjecture.
Since the cause is likely to be within POE's implementation, we have
room for improvement. For example, it's been suggested on
irc.perl.org #poe to make POE's I/O events latency tunable. An
interactive application might choose better responsiveness, while a
bulk transfer application (such as yours?) could be tuned for higher
throughput. It's an interesting idea, but I don't have resources to
pursue it. If someone else does, I'll be happy to provide advice and
guidance, here, in private e-mail, or on IRC.
--
Rocco Caputo - [email protected]
On Mar 23, 2009, at 07:18, howard chen wrote:
I have 2 simple scripts which connect to NNTP server to fetch 100
articles (no writing to local disk) in order to test the performance
of two frameworks.
I have found Twisted is faster by at least 2 times which make me
surprised.
I know it is not fair to compare two frameworks by this simple test.
But I think 2 times is quite much.
Therefore, I upload the code online to see if any experts can find
any problem?
Thanks.
http://howachen.googlepages.com/nntp.py
http://howachen.googlepages.com/nntp.pl
time python nntp.py news.microsoft.com microsoft.public.access
access.mbox
18 sec
time perl nntp.pl
37 sec
Any comments?
