Slow konqe on a PXA250

[Paul Chitescu]

> Hi all friendly people!
>
> For evaluation purposes, I'm currently running Linux on an Intel XScale
> PXA250 Processor (on one of them Intel development boards), running at
300MHz
> with the memory clocked at 100MHz (or at times at 50MHz since the boards
are

Tell me about it - our SA-1110 system never manages to decompress the kernel
with a SDRAM clock over 90MHz :(

> not really working at 100MHz very often). I have a 640x480 pixels LCD
using
> 8bit colour. (Uh, fyi, the XScale is an ARM thing.)

SA-1110 being its precursor.

> I have konqueror-embedded-snapshot-20020311 built and running on top of
the
> Qt/E 2.3.2. (I have not done any particular optmizations or
configurations,
> merely got them working using the smallest amount of effort possible.) I
run
> the browser full-screen.
>
> The system has 64MB ram and 10Mbit ethernet connection.

Less RAM for me but it doesn't matter. Missing a swap if the system rans out
of memory the kernel starts killing processes and the first candidates are
the KIOs - easy to spot in the logfile.

Yes, memory leaks big time but it's Javascript related. Normally there are
few or no global objects on page. And it takes a while until oom_kill shows
up.

> I find konqueror running unexpectedly slow on this configuration, and I'll
> share some test observations:
>
> I setup a single 100-bytes something HTML page, using only four images
(each
> being 600x100 pixels). No text.

I have something similar, 4 gradient bars (to verify the order of bits,
uh-oh) of 600x100 each. From a local CompactFlash these load in under 1.5s
and from a local web server in about 2s. Note that these are PNG images with
a very small size, under 1K each.

> I enter the URL for that HTML file in konqueror and press return. I can
see
> in my ethereal trace that the fetch of the HTML file flies by really fast.
> (The http server is connected on the same hub as the devboard.)

Similar test configuration. The nameserver is connected to the same hub.
Pinging from the device shows about 1.5ms (0.2ms for a K7/1400)

> Then it can take up to almost 3 seconds of no network traffic at all
before
> konqueror decides to get the first image. It then gets the second and
third
> images within a quite reasonable time (fractions of a second), just to
again
> make a pause before the 4th image is retrieved. This behavior is easy to
> repeat and seem to be roughly the same whenever I test this.
>
> So, having 4 JPEG images at ~7-8 KB each, takes roughly 7-8 seconds to
> display! When having done that, and I enter the address again I get the
> images from the internal cache, but that *too* takes many seconds to
display!
>
> Changing the JPEG images into GIF made the images ~60KB each but the
timings
> when loading from the net didn't differ much, but getting the images from
the
> cache is *a lot* faster than in the JPEG case!!

Jpeg is more computing-intensive than other formats so the result is normal.
Changing the JPEG images to pre-dithered PNG versions improves noticeably
the speed.

> I've also produced a text file using 30 K text (the GPL license) and
> displaying this here takes roughly 12-15 seconds.

Wow! It takes for me about 2.5s when the text/plain is downloaded from a
server and htmlized by a filter (implemented in sed) although the layout of
<pre> should be easy.

> All these times seem very excessive. I can't understand how it can be this
> slow.  Can anyone bring me any clues on what I might be doing wrong or
what I
> can do next to improve performance on my setup?

Try to place the media on a local storage so you can check if it's network
or computing related. See what's the BogoMips rating measured by the kernel
at boot time.

Maybe you also tell us what kernel, toolchain and jpeg library you use.


Paul Chitescu

--
 Paul Chitescu
 pchitescu@null.ro   http://pchitescu.null.ro/   ICQ:22641673
 Any spammers will be painfully squeezed into /dev/null