[sugar] On Matploblib for the XO

[Ivan Krstić]

Hi Fernando, John,

I wanted to restart the discussion around Matplotlib, IPython, and the  
entire scientific/mathematical environment idea on OLPC laptops. I  
responded to the mail below previously, but I'm not sure what happened  
after that.

Arjun Sarwal, who is the author of the Measure activity that I  
demonstrated at SciPy, has recently been looking into plotting  
libraries for the XO, and (understandably) Matplotlib caught his eye.  
Arjun and others -- I'm forwarding Fernando's original e-mail below  
for your reference.

I am in full support of the Matplotlib/sympy/IPython combo getting up  
and running on the XO, on the way to an awesome complete sci/math  
environment. I hope Fernando and John will still consider giving us a  
hand.

Thanks!

Cheers,
Ivan.



On Aug 19, 2007, at 10:38 PM, Fernando Perez wrote:
> In order to estimate the sizes one could expect to see with a
> matplotlib build for the OLPC, I tried to build it with only the GTK
> Cairo backend (though AGG still needs to be built for internal use,
> since it's everywhere) and then added GTKAgg as well.  These were the
> sizes I got:
>
> tlon[build]> d *.tgz
> /home/installers/src/scipy/matplotlib/build
> -rw-r--r-- 1 fperez 4674627 2007-08-19 19:43 mpl_cairo_only.tgz
> -rw-r--r-- 1 fperez 5555121 2007-08-19 20:25 mpl_with_gtkagg.tgz
>
> However, the Cairo backend is still not very mature and doesn't
> perform very well; in particular it lacks proper transparency support,
> so using GTK+Agg might be a better alternative.  These sizes are what
> I got from doing a build and quickly removing some obvious things (as
> well as deactivating support for Qt, Tk and WX).  It's quite likely
> that John might suggest how to trim it further, I was afraid of
> cutting things I shouldn't so I went in conservatively.
>
> John will correct me if I'm wrong, but I think that since you already
> ship a real GTK, the GTKAgg backend should work just fine.  Agg is
> very fast, has extremely high quality, and the GTKAgg backend is
> probably the most mature and widely tested one of all.  If 5.5 MB is
> an acceptable size for you, here are the numbers for the other pieces
> of the puzzle.
>
> For IPython:
>
> -rw-r--r-- 1 fperez 713456 2007-08-19 21:18  
> ipython_installed_nodocs.tgz
> -rw-r--r-- 1 fperez 808717 2007-08-19 21:16 ipython_installed.tgz
>
> IPython comes in at a bit under 1MB, keeping the source and HTML
> manual but removing the PDF docs.  Removing the HTML manual saves
> about 100k.
>
> As for sympy, I just installed it and tarred it:
>
> -rw-r--r--  1 fperez 1686428 2007-08-19 21:21 sympy_installed.tgz
>
> I'm not familiar enough with it to know what can be cut out, though I
> know their 3d plotting is OpenGL based, so that can probably go.
> You'd need to get the sympy team involved, though Ondrej sounded
> enthusiastic when I spoke with him.
>
>
> So it looks like the ipython/matplotlib/sympy combo would cost you
> around 8MB total.  That's not counting numpy, but you told me you'd
> already decided to include that for other reasons (matplotlib can't
> run without numpy, nor can one do many key numerical operations
> without it at reasonable speeds).
>
> One final question: have you run any numpy code on the olpc to get an
> idea for its floating-point performance?  Here are a couple of
> reference points and an easy way to get some timing info:
>
> On my 5 1/2 year old laptop, a Pentium-III at 1.13 GHz with only SSE
> (no SSE2) instructions, a typical test time for numpy is:
>
> maqroll[~]> time python -c "import numpy;numpy.test(10)" > & /dev/null
> 3.484u 0.196s 0:03.73 98.3%     0+0k 0+0io 0pf+0w
>
> On my desktop, a dual-core Athlon 4400 (2.2GHz clock):
>
> tlon[~/tmp]> time python -c "import numpy;numpy.test(10)" > & /dev/ 
> null
> 1.420u 0.144s 0:01.56 100.0%    0+0k 0+0io 0pf+0w
>
>
> My laptop is perfectly usable as a matplotlib/numpy computing
> environment at these speed, and you could probably take a factor of 4
> hit on performance and still be OK for the numerics.  The graphics may
> get a little unpleasant if your box is much more than 2-4 times slower
> than my laptop, I suspect, but it might be OK. Also, I think that if
> one disables Agg and uses only the plain GTK backend, things go faster
> (John will correct me if I'm wrong).
>
>
> In any case, I hope you find this useful.  Feel free to forward it to
> those on your team you deem appropriate.

--
Ivan Krstić <krstic at solarsail.hcs.harvard.edu> | http://radian.org