Not Even My Wife Reads This

I'm back from PyCon in Silicon Valley. (If you don't know what PyCon is, it's 2000+ nerds talking about the Python programming language for a few days.) Here are some highlights I remember, in case you're looking for recommendations of videos to watch. (Of course, it's fantastic that they record everything and make it all available for free, here )

The introduction featuring dancing robots was impressive. I remember that just a few years ago it was hard to teach robots to walk, and now you (or your rich friend) can buy a robot that can be programmed to dance (among many other things) for a few thousand bucks. They ended up giving away one of the robots at the end, a pretty nice prize.

Paul Graham's keynote about ten really-hard-but-not-quite-impossible startup ideas was awesome, probably the best thing I saw at the conference.

Stormy Peters' keynote about freedom and privacy and stuff was okay; I agreed with her about pretty much everything, but I don't know if someone who didn't would actually have been swayed.

Guido van Rossum's keynote about "trolls" who whine about Python didn't do much for me.

Katie Cunningham's intro to hosting a web site in the cloud and using Fabric to maintain it was very good. I've never used Fabric but I plan to start.

Raymond Hettinger's talk about subclassing was predictably excellent. It was a very balanced talk about when it makes sense to subclass (because it lets you write less code) and when it doesn't, not the typical object oriented programming 101 nonsense where they explain how to subclass but not why it's not always the right thing to do. But since I pretty much already agreed with everything he said, I don't think I got anything out of it.

Larry Hastings' low-level stepping through Python was fun for a very select audience. Most people probably don't need that level of detail.

Paul McMillan's talk about security was very good, and useful for any web programmer who doesn't know this stuff.

Chris McDonough's talk about PDB (the Python debugger) was pretty good. I picked up a few tricks I didn't know, even though I've used PDB a bunch of times.

Moshe Zadka's talk about writing resilient programs that recover from crashes was great. He mentioned several techniques I hadn't even thought of, let alone used.

Geremy Condra's talk about security testing was good, but not quite as mind-blowing for me as his previous talk about timing attacks.

Glyph Lefkowitz's talk about low-level networking (and not Twisted) was great, though I already knew pretty much everything he covered so I probably should have been elsewhere.

The PyPy survey talk by Maciej Fijalkowski, Alex Gaynor, and Armin Rigo was actually more like two talks. Maciej and Alex gave kind of a general introduction to PyPy, then Armin talked about software transactional memory, making an analogy to garbage collection. I already follow PyPy so I didn't get too much out of this talk, but it's worth watching if you don't.

Benjamin Peterson's talk about the PyPy just-in-time compiler was more in-depth and thus more interesting to me. I'd heard most of the material before but it's complicated enough that I needed to hear it again.

Michael Bayer's talk about SQLAlchemy was okay. It was more advocacy of the right way to do a flexible database toolkit (make it flexible) than practical advice on how to use SQLAlchemy.

Both Geographic Information System talks I attended were excellent. Jason Scheirer talked about GIS and maps at a more theoretical level, covering the basic terminology and map projections. Zain Memon's talk was more practical, basically how to use a bunch of libraries to quickly create a map-oriented web site. I need to do some GIS stuff at work, so these two talks definitely justified the money my employer spent to send me to the conference.

Jacqueline Kazil and Dana Bauer's talk about mining public data released by local governments was really good. I wasn't at all familiar with most of the tools they used, so I'll need to watch the video and take notes.

Chris Lambacher's talk about cross-compiling Python and C extensions for embedded systems covered something else I need to do at work, so I was really hoping it would solve all my problems. Unfortunately, his slides showed a big mess of low-level hacks around tools that aren't really designed for cross-compiling, rather than the magic bullet I wanted (but didn't expect to find.) It's a hard problem, but maybe the emergence of ARM devices will make cross-compiling more mainstream and the tools will improve.

Erik Rose's talk about parsing MediaWiki files (which are too complicated to parse with simple regular expressions) gave us an entertaining survey of Python parsing libraries. He covered a handful in some detail, and has a survey of lots and lots more on his web site.

Ryan Kelly's talk about frozen standalone programs didn't really go much into running programs like py2exe, but rather dealt with secondary issues like auto-updaters, binary compatibility with old operating system versions, and code signing. It was interesting, but didn't solve my short-term need to make PyInstaller and PyGTK cooperate on 64-bit MacOS.

Brandon Rhodes' talk about linkers and virtual memory was really excellent, but unfortunately covered things I'd already learned, so I should have been somewhere else.

The lightning talks were a mixed bag. I enjoyed many of them, but certainly don't remember enough to review them all. I'd kind of like to attend a conference that was nothing but lightning talks, five minutes each on hundreds of subjects.

The poster session was good. There were about 30 posters, with their authors handy to talk, and about half of them covered subjects that interested me at least a bit.

The Expo Hall showed that the tech economy is doing fine, even if the larger economy is still kind of bleak. There were dozens of sponsors with booths, and at least half of them were desperately trying to hire people. If you're an out-of-work Python programmer, definitely go to PyCon.

Finally, I spent a day and a half at the sprints, hacking on PyPy. I fixed 3 or 4 minor bugs and learned a bit in the process. I would love to have spent another couple of days at the sprint, but work and family called. (I had actually planned to spend a couple of days playing tourist rather than coding, but changed my mind. So I'll have to go back and look at redwoods some other time.)

In some ways, this was the best PyCon I'd attended since the very first one in DC. The WiFi worked really well for me for a change, even though I only brought a cheap netbook. I managed to avoid attending any really bad talks. I got to write some useful code and talk to a lot of people and even play in a free poker tournament. And I caught not one but two door prize frisbees. (Unfortunately, I only won a couple of books, not a robot or an iPad, but maybe next year.)

I reran my Project Euler benchmark with PyPy 1.8, PyPy 1.7, Psyco 1.6, and CPython 2.7.2. (On Gentoo Linux x86.)

Total runtime for 127 small programs was 184.18s for PyPy 1.8, 185.55s for PyPy 1.7, 343.51s for Psyco, and 533.09s for CPython.

Or in relative terms, PyPy 1.8 was less than 1% faster than 1.7. Both versions of PyPy were almost twice as fast as Psyco, and almost three times as fast as CPython.

Full results follow:

PyPy 1.6 was just released, so I ran my Project Euler benchmark script again.

The script runs my pile of Project Euler solver programs across various implementations of Python available on my computer, and gives results only for the ones that finish successfully in less than a minute on all of them.

The contenders this time were PyPy 1.6, CPython 2.7.2, CPython 2.6.7 with Psyco, and Jython 2.5.1+. (Unladen Swallow appears to be dead, and IronPython is surprisingly painful to install on Linux, so I left them out.)

This time, 125 programs qualified, out of 197 working solvers. (Some programs require the gmpy library, which doesn't currently work on PyPy or Jython. And some use Python 2.7 syntax, which won't work with Psyco or Jython. And some are just too slow.)

Once again, PyPy is fastest, followed by Psyco, then CPython, then Jython. PyPy and Psyco are roughly twice as fast as CPython, and Jython is roughly twice as slow as CPython. (Though a lot of that is startup overhead; Jython gets more competitive for longer-lived programs.)

In a previous post, I mentioned that PyPy was the fastest Python implementation for most of my Project Euler programs, but that it was very slow for a few of them.

This is no longer the case. The jit-generator branch was merged a few days ago, fixing a weakness with code that uses generators. And now PyPy is clearly the fastest Python implementation for this code, with both the most wins and the lowest overall time. Psyco is still pretty close. Both are a bit more than twice as fast as CPython.

I compared PyPy trunk, Unladen Swallow trunk, Jython 2.5.1+, CPython 2.6.6 with Psyco, and CPython 2.7.

PyPy is very strong across the board. Its worst result is on euler94, a Sudoku solver that heavily uses sets and copy.deepcopy.

Psyco still does very well, but it doesn't work on Python 2.7 yet and still doesn't work on amd64, so it feels more and more like a dead end.

Unladen Swallow hasn't had a commit since August. I suspect it's just resting, not dead, but it's falling behind PyPy in performance. Version 2.8 of LLVM has been released, but Unladen still requires version 2.7.

CPython is the baseline. I used 2.7, the latest Python 2. (Some of my Euler programs work in Python 3; some don't.)

Jython is by far the slowest. Its large startup overhead hurts on the easy problems, and the mature HotSpot JIT isn't enough for it to catch up on the harder ones. Jython does have the advantage of being free-threaded, but since this code was originally written for CPython it rarely uses multiple threads and so doesn't really benefit.

Here are the numbers.

At PyCon, Maciej from the PyPy team asked for programs that the PyPy JIT did a really bad job on.

I have a directory full of over 200 little Python programs that attempt to solve Project Euler math problems. Only 186 of them get the right answer, and fewer of them run in a reasonable period of time. But this led me to concoct a quick benchmark comparing the Python implementations installed on my computer: CPython, CPython with Psyco, Unladen Swallow, and PyPy. (I haven't tested them with Jython or IronPython yet. Maybe someday.)

Basically, I just run all of euler*.py under four different Python implementations, record how long it took each implementation to run each program, and throw out the results for any programs that didn't work in all four versions of Python, or that didn't finish in 60 seconds or less on all of them.

Here are the results, for recent svn checkouts of PyPy and Unladen, CPython 2.6.4, and CPython 2.6.4 plus Psyco 1.6. (Don't get too obsessed with the exact results, since it's just a private benchmark and most of the code is unpublished. This is just background for the real information at the bottom of the post.)

So all the JITs are faster than CPython for this set of programs, but none is twice as fast. PyPy has the most wins, but Psyco has the lowest total time.

While PyPy does very well overall, it's an order of magnitude slower than all the others on 3 of the programs: 14, 135, and 174. If PyPy fixed whatever made those ones slow, it would definitely be the overall leader.

Here's my euler14.py

#!/usr/bin/env python

"""Project Euler, problem 14

The following iterative sequence is defined for the set of positive integers:

n -> n / 2 (n is even)
n -> 3n + 1 (n is odd)

Which starting number, under one million, produces the longest chain?
"""

def next_num(num):
    if num & 1:
        return 3 * num + 1
    else:
        return num // 2

MAX_NUM = 1000000

lengths = {1: 0}

def series_length(num):
    global lengths
    if num in lengths:
        return lengths[num]
    else:
        num2 = next_num(num)
        result = 1 + series_length(num2)
        lengths[num] = result
        return result

def main():
    num_with_max_length = 1
    max_length = 0
    for ii in range(1, MAX_NUM):
        length = series_length(ii)
        if length > max_length:
            max_length = length
            num_with_max_length = ii
    print num_with_max_length, max_length

if __name__ == "__main__":
    main()

Why is 14 so slow? My guess would be that it's because it heavily mutates a global dictionary.

Here's euler135.py:

#!/usr/bin/env python

"""Project Euler, problem 135

Given the positive integers, x, y, and z, are consecutive terms of an
arithmetic progression, the least value of the positive integer, n, for which
the equation, x**2 - y**2 - z**2 = n, has exactly two solutions is n = 27:

34**2 - 27**2 - 20**2 = 12**2 - 9**2 - 6**2 = 27

It turns out that n = 1155 is the least value which has exactly ten solutions.

How many values of n less than one million have exactly ten distinct solutions?
"""

"""
x**2 - y**2 - z**2 = n
y = z + d
x = z + 2 * d
d > 0
z > 0
(z + 2 d) ** 2 - (z + d) ** 2 - z ** 2 = n
(z**2 + 4dz + 4d**2) - (z**2 + 2dz + d**2) - z**2 = n
z**2 + 4dz + 4d**2 -z**2 - 2dz - d**2 - z**2 = n
-z**2 + 2dz + 3d**2 - n = 0
a = -1
b = 2d
c = 3d**2 - n
z = (-b +/- sqrt(b**2-4ac)) / 2a
z = (-2d +/- sqrt(4d**2+4(3d**2-n))) / -2
z = (-2d +/- sqrt(4d**2+12d**2-4n))) / -2
z = (-2d +/- sqrt(16d**2-4n))) / -2
z = (-2d +/- 2 sqrt(4d**2 - n)) / -2
z = (-d +/-  sqrt(4d**2 - n)) / -1
z = d +/- sqrt(4d**2 - n)
4d**2 - n >= 0
n <= 4d**2
4d**2 - n is a perfect square
"""

def main():
    limit = 1000000
    counts = limit * [0]
    for d in range(1, limit / 4 + 1):
        d24 = d ** 2 * 4
        if d24 < limit:
            start = 1
            counts[d24] += 1
        else:
            start = int((d24 - limit) ** 0.5)
        if start < 1:
            start = 1
        for root in xrange(start, 2 * d):
            n = d24 - root ** 2
            if n <= 0:
                break
            if n < limit:
                z = root + d
                y = z + d
                x = y + d
                #print "n", n, "d", d, "root", root, "x", x, "y", y, "z", z
                #assert x**2 - y**2 - z**2 == n
                counts[n] += 1
                if d > root:
                    z = d - root
                    y = z + d
                    x = y + d
                    #print "n", n, "d", d, "root", root, "x", x, "y", y, "z", z
                    #assert x**2 - y**2 - z**2 == n
                    counts[n] += 1
    total = 0
    assert counts[0] == 0
    for val in counts:
        if val == 10:
            total += 1
    print total

if __name__ == "__main__":
    main()

Why is PyPy so slow on this one? Well, maybe it's because all the work happens in one function, if PyPy doesn't JIT a function while it's still running. Or maybe it's the list with a million elements in it.

Finally, euler174.py:

#!/usr/bin/env python

"""Project Euler, problem 174

We shall define a square lamina to be a square outline with a square "hole" so
that the shape possesses vertical and horizontal symmetry.

Given eight tiles it is possible to form a lamina in only one way: 3x3 square
with a 1x1 hole in the middle. However, using thirty-two tiles it is possible
to form two distinct laminae.

If t represents the number of tiles used, we shall say that t = 8 is type L(1)
and t = 32 is type L(2).

Let N(n) be the number of t <= 1000000 such that t is type L(n); for example,
N(15) = 832.

What is sum(N(n)) for 1 <= n <= 10?
"""

from math import ceil
from collections import defaultdict

def gen_laminae(limit):
    """Yield laminae with up to limit squares, as tuples
    (outer, inner, squares)"""
    for outer in range(3, limit // 4 + 2):
        if outer & 1:
            min_min_inner = 1
        else:
            min_min_inner = 2
        min_inner_squared = outer ** 2 - limit
        if min_inner_squared < 0:
            min_inner = min_min_inner
        else:
            min_inner = max(min_min_inner, int(ceil(min_inner_squared ** 0.5)))
            if outer & 1 != min_inner & 1:
                min_inner += 1
        outer_squared = outer ** 2
        for inner in range(min_inner, outer - 1, 2):
            squares = outer_squared - inner ** 2
            yield (outer, inner, squares)

def main():
    squares_to_count = defaultdict(int)
    for (outer, inner, squares) in gen_laminae(1000000):
        squares_to_count[squares] += 1
    histogram = defaultdict(int)
    for val in squares_to_count.values():
        if val <= 10:
            histogram[val] += 1
    print sum(histogram.values())

if __name__ == "__main__":
    main()

Perhaps this one is slow because it heavily uses generators and a defaultdict?

Anyway, I'm not claiming any of these are great programs. I know others have much better solutions to these problems. And this code is certainly not optimized for PyPy. They're just examples of small programs that were originally written for CPython, and do work correctly on PyPy, but happen to be much slower on PyPy than on CPython.

Again, note these are the exceptions not the rule. PyPy runs most pure-Python code great. Just not these three programs.

PyCon was in Atlanta this year, which got me away from the record snow in the DC area.  Yay for winter conferences being in warm places.  (Nothing against Chicago, but Chicago conferences should be in the summer.)

Attendance was about 1100, a little more than last year, but nothing like the crazy growth PyCon saw when the economy was strong. I think that Python is still growing but travel and recruiting budgets are down.

The best talk I attended was Raymond Hettinger's.  It was about using the right container classes to solve computationally expensive problems.  He pointed out that every ordered dict implementation except the one he just added to Python 3.1 had O(n) deletes due to using a list for storing the order, while his had O(1) deletes because he used a linked list with a dict index.  I hung my head in shame because I've written odict twice (once at a former employer, again from scratch for Slugathon) and done it "wrong" both times.

The most useful thing I attended was the Twisted Open Space.   Because my employer really wants IPv6 in Twisted, and I've submitted a patch but not had it accepted because of reverse compatibility concerns.  Actually talking to most of the core Twisted team at once in person really helped clarify what we need to do to break the logjam and get the patch moving forward again.  That ten minutes probably justified the cost of sending me to PyCon this year.  (And I wish I could go to the sprints and maybe actually get this change into Twisted this week, but I can't.  Next year I really want to go to at least a couple of sprint days.)

Rackspace Cloud was there as a sponsor with a booth, which reminded me that I should have blogged about my experience using Rackspace Cloud.  Basically, they stood up a virtual small Ubuntu (they have other choices too) server for me in about 5 minutes, for about $12 per month (plus bandwidth, more money for more memory), and it just stinking worked.  I've since turned it off because Slugathon isn't done yet so I don't really need a dedicated game server yet, but I'll definitely be back when it is.  The only negative is that there's no way to setup a cap at which the server turns itself off, so if you get Slashdotted or DOS attacked you may get a high bandwidth bill.

Negatives:

Airline travel sucks.  You already know this.

Guido's keynote was just taking questions via Twitter, and the signal-to-noise ratio was awful.  (Yes, I'm an old Angry Unix Guy.  Get off my lawn and take your Twitter and your Facebook and your iPhone with you.)  If Guido doesn't want to do a real keynote, that's fine; why not let someone else have the slot?

I'm no longer in favor of invited talks, because one of them was just content-free pattern metababble that never would have been accepted if the speaker had had to do a proposal.  (But, in fairness, all the other invited talks I attended were excellent.)

There were several talks that I really enjoyed and thought were great fun, but where I didn't really learn anything.  So they validated my existing opinions but didn't stretch my brain at all.  (Larry Hastings actually pointed out before his talk about micro-optimizations that it was just nerd porn and that it would be entertaining but nobody would learn much.  I applaud him for his honesty.)  I guess that's natural when you've been doing something for a long time and have attended the same conference a bunch of times.  I need to attend fewer talks and spend more time just talking to people.

The board game social wasn't as awesome as last year because there wasn't a huge pile of games to pick from.  (I think last year a game store donated some games, and you can't expect that to happen every year.)  So I will bring at least one game next year.

My 7-year-old laptop lost its WiFi connection whenever things got crowded.  The networking people do their best, but 1000 laptops crammed into a small area means the newer stronger ruder WiFi cards will crowd out the older weaker more-polite ones, no matter how many access points there are.  I plugged into a switch when I could, and lived without WiFi when I couldn't.  My laptop is heavy and has poor battery life anyway, so I think it's time to retire it in favor of a netbook.

I've been working on Slugathon on-and-off (but mostly off) since 2003. The bare minimum feature set that I consider necessary to do a serious alpha release is almost done. (Release early, release often; but if you release before you actually do anything useful you might just leave a bad first impression and scare off your potential users.) So I've been thinking about installers, and portability to MacOS and Windows, and whether my current hosting solution (svn and Trac provided for free to open source Python projects by WebFaction; thanks WebFaction) would cut it when I had users and possibly more contributors.

I like Trac, a lot. I now strongly prefer Git to Subversion. You can use Git and Trac together with the GitTracPlugin, but it's hard to find a host that has it installed. (SourceForge, for example, now offers both Git and Trac, but not together.) So I decided to go with github. They're the most popular Git host, their paid hosting business seems to be solid enough that they're unlikely to go away anytime soon, they offer free hosting for open source projects, and they have a wiki and an issue tracker.

So this morning I experimentally moved everything from Subversion+Trac to github. I've administered both Svn/Trac and Git/Trac installations at work, so I knew how to do this on a local machine where I had control, but doing things locally is different from doing things on a remote hosting service.

Here's what I did on github:

1. Signed up for an account. (I've poked around github plenty of times, and cloned other people's repositories there to my local box, but I never needed to register until I actually wanted to move my own repository there. Which is exactly the way it should work.)

2. Uploaded an ssh key. I already know ssh so it was just a matter of posting ~/.ssh/id_rsa.pub into a web form.

3. Re-cloned my Subversion repository to my local box, using git-svn, using the –no-metadata option to remove all the ugly "git-svn-id" blobs in the log. I'd done this before when we switched from Subversion to Git at work, but I didn't remember the exact syntax. Luckily it was right there on Github's help page so I didn't even need to check a Git man page.

4. Uploaded my new local Git repository to Github.

5. Realized that I'd forgotten to use the –authors-file option to convert the author metadata from Subversion format (bare username) to Git format (firstname lastname email). Oops, this was also on github's help page but I'd been in too much of a hurry. I poked around github until I found out how to delete a repository, deleted the one I'd just made, redid the git-svn clone with the missing option, and re-uploaded my repository.

6. Basically cut-and-pasted my wiki pages (there were only 8 of them) from Trac to Github's wiki. I did a little tweaking to convert formatting from Trac format to the Textile format that Github uses, but didn't obsess about getting every little format perfect. One thing that github's wiki doesn't support well is images; they let you link to external images, but not post attachments directly into the wiki. So for now the thumbnail screenshots are there, but they don't link to the larger versions.

7. Manually ported all my Trac tickets to github issues. I had 62 tickets, 30 closed and 32 open. Both Trac and the github issue tracker share the same #number format for referring to issues, and I mention ticket numbers inside commit messages, so I thought it was necessary to add all the already-closed issues for historical reasons. I also changed any references to Subversion commit numbers to Git commit ids inside the issues. This was an annoying data entry task, but with only 62 tickets it was faster to just do it by hand than to write some awesome Trac to github conversion script. (Which would require either cooperation from github, or advanced web scraping, since github is a rather fancy web site with JavaScript everywhere.)

8. Changed Slugathon's SourceForge page to point at github rather than WebFaction. (I'm keeping a SourceForge presence for the project because they offer useful things that Github does not, like project mailing lists. Also, if SourceForge ever integrates Git and Trac and I become dissatisfied with Github's new and fairly minimal issue tracker, I might move everything to SourceForge.)

9. Changed the main wiki page at WebFaction to point to github, too. (I will eventually ask WebFaction to delete the project, but first I want to make sure that the move was a good idea. And give Google enough time to cache the version of the page that shows the redirect to github, so people don't think the project just disappeared.)

10. Created a static web page using the instructions at pages.github.com. This feels somewhat redundant with the github wiki, but it gave me a place to park my images.

My conclusion so far: I still like Git a lot more than Subversion. I like github's wiki and issue tracker less than Trac, but they seem to be good enough.

I have a TrendNet TE100-P1P print server. It's a little $30 gizmo with an Ethernet jack on one end and a parallel printer port on the other, so that you can turn a printer into a network printer without hanging it off a computer. Bought it because my new motherboard didn't have a parallel port. The other benefit is that now the printer works when my main desktop computer is off.

Anyway, we had a power failure, and the printer stopped working. No "lexmark" found in local DNS. No DHCP lease for the TrendNet gizmo.

The gizmo was still there, but I couldn't find it. There's no UI on the device, just the network jack. So I wrote a little script (below) to portscan my LAN for anything that answered on port 631. That found it.

Turns out that when the power goes out, the print server reconfigures its network settings from DHCP to fixed. It keeps its last network settings, but the DHCP / DNS server no longer has any record of them, so you can't find it unless you remember what its last IP was.

Here's the script. Because I had 700+ addresses to check, and it sometimes takes a few seconds for a failed socket connection to timeout, I used 250 threads to make it faster. (I initally tried using a separate thread for each IP, but I hit an OS limit at 255 threads, and decided to use a fixed-size pool and two shared work queues.)

#!/usr/bin/env python

"""Hunting for the TrendNet print server"""

import socket
import threading
import Queue

socket.setdefaulttimeout(5)

port = 631
num_threads = 250

class ConnectThread(threading.Thread):
    def __init__(self, in_queue, out_queue):
        threading.Thread.__init__(self)
        self.in_queue = in_queue
        self.out_queue = out_queue
        self.setDaemon(True)

    def run(self):
        while True:
            ip = self.in_queue.get()
            try:
                sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
                sock.connect((ip, port))
            except socket.error:
                self.out_queue.put((ip, False))
            else:
                self.out_queue.put((ip, True))
            self.in_queue.task_done()

def main():
    job_queue = Queue.Queue()
    result_queue = Queue.Queue()
    count = 0
    for net in range(3):
        for addr in range(1, 255 + 1):
            ip = "192.168.%d.%d" % (net, addr)
            count += 1
            job_queue.put(ip)
    for unused in xrange(num_threads):
        ConnectThread(job_queue, result_queue).start()
    job_queue.join()

    for unused in xrange(count):
        ip, status = result_queue.get()
        if status:
            print ip

if __name__ == "__main__":
    main()

Someone on the PyGTK mailing list just asked which is faster for drawing rectangles, GDK or Cairo.

I wasn't sure, so I wrote a silly little test program.  Note that it's not quite apples-to-apples as the Cairo rectangles have variable transparency while the GDK rectangles are opaque.

On my box, running Gentoo Linux and the latest stable versions of everything, the Cairo version draws 500 rectangles in about 0.01 to 0.03 seconds, while the GDK version takes about 0.13 to 0.14 seconds.  So Cairo is faster.

To run this, just cut-and-paste into an editor window, save the file as rectangles.py, and run "python rectangles.py"

#!/usr/bin/env python

"""GTK rectangle drawing speed test, GDK vs. Cairo

David Ripton 2008-12-03
MIT license
"""

import time
import random

import gtk

NUM_RECTS = 500

handle_id = None

def main():
    window = gtk.Window()
    window.connect("destroy", gtk.main_quit)
    window.set_default_size(800, 600)
    vbox = gtk.VBox()
    window.add(vbox)
    area = gtk.DrawingArea()
    vbox.pack_start(area)
    gdk_button = gtk.Button("GDK")
    gdk_button.connect("clicked", on_gdk_button_clicked, area)
    vbox.pack_start(gdk_button, expand=False)
    cairo_button = gtk.Button("Cairo")
    cairo_button.connect("clicked", on_cairo_button_clicked, area)
    vbox.pack_start(cairo_button, expand=False)
    window.show_all()
    gtk.main()

def on_gdk_button_clicked(button, area):
    global handle_id
    if handle_id is not None:
        area.disconnect(handle_id)
    handle_id = area.connect("expose-event", on_area_exposed_gdk)
    area.queue_draw()

def on_cairo_button_clicked(button, area):
    global handle_id
    if handle_id is not None:
        area.disconnect(handle_id)
    handle_id = area.connect("expose-event", on_area_exposed_cairo)
    area.queue_draw()

def on_area_exposed_gdk(area, event):
    t0 = time.time()
    width, height = area.window.get_size()
    colormap = area.get_colormap()
    gc = area.get_style().fg_gc[gtk.STATE_NORMAL]
    for ii in xrange(NUM_RECTS):
        r = random.randrange(0, 65535 + 1)
        g = random.randrange(0, 65535 + 1)
        b = random.randrange(0, 65535 + 1)
        gc.foreground = colormap.alloc_color(r, g, b)
        x = random.randrange(0, width)
        y = random.randrange(0, height)
        w = random.randrange(0, width - x)
        h = random.randrange(0, height - y)
        area.window.draw_rectangle(gc, True, x, y, w, h)
    t1 = time.time()
    print "gdk drew %d rectangles in %f seconds" % (NUM_RECTS, t1-t0)

def on_area_exposed_cairo(area, event):
    t0 = time.time()
    cr = area.window.cairo_create()
    width, height = area.window.get_size()
    for ii in xrange(NUM_RECTS):
        r = random.random()
        g = random.random()
        b = random.random()
        a = random.random()
        cr.set_source_rgba(r, g, b, a)
        x = random.randrange(0, width)
        y = random.randrange(0, height)
        w = random.randrange(0, width - x)
        h = random.randrange(0, height - y)
        cr.rectangle(x, y, w, h)
        cr.fill()
    t1 = time.time()
    print "cairo drew %d rectangles in %f seconds" % (NUM_RECTS, t1-t0)

if __name__ == "__main__":
    main()

Rant: It is way too hard to post code with Wordpress. First I tried the "code" button, but all my indentation was destroyed (bad for any code, fatal for Python). Then I tried the "b-quote" button; same effect. Then I switched to HTML mode and hand-inserted a "pre" tag, which preserved the indentation. But Wordpress then proceded to vandalize my code with "smart" quotes. (Have you ever seen anything with "smart" in its name that actually was?) Luckily it was simple to find and install the wpuntexturize plugin, a few lines of PHP that eradicate Moron Quotes. But why the hell are they there in the first place, let alone enabled by default, let alone enabled by default inside a "pre" tag?

Stephen Waterbury added a WxPython client to my AmpChat example program. (Mentioned earlier in this post.)

His Mercurial repository is here.

This is still just example code, but now it's an example of multiple things:

• how to use Twisted AMP
• how to use Twisted with PyGTK (trivial because gtk2reactor rocks)
• how to make WxPython's mainloop run in parallel with Twisted's using threads
• how to write the same simple GUI program in both PyGTK and WxPython
• how to use Mercurial to contribute to a project without commit rights in the original repo

Anyone want to contribute a PyQT or Tk version?