Another thing that seems to be ignored here is how well gzip performs when embedding multiple images in a single stylesheet. When we implemented this feature we were skeptical about possible side-effects of say, embedding the same image multiple times, or embedding images that might not end up being used. Our tests showed that gzip was so aggressive at combining common image headers and in some cases color pallets or actual pixel data, that we would have to have a pretty extreme situation for these sorts of things to actually cause performance issues.

- Trevor


On Wed, Sep 25, 2013 at 10:53 PM, Jon Robson <jdlrobson@gmail.com> wrote:
Yup very flawed but interesting never the less. Putting into the back
of the memory bank. Thanks for sharing Ori and thanks Brion for
pointing out the flaws that I also skimmed over :)

On Mon, Sep 23, 2013 at 5:26 PM, Matthew Flaschen
<mflaschen@wikimedia.org> wrote:
> On 09/23/2013 06:11 PM, Brion Vibber wrote:
>>
>> Per comments on the article, I wouldn't read much into it -- the
>> benchmark is pretty flawed. It doesn't measure anything representative
>> of typical usage.
>
>
> Yeah.  They inexplicably *only* used a single size, a 17.8kB PNG (18227
> bytes).  If you're not sure why that's a problem, check the size of some of
> the images MW embeds.  Examples from a quick grep:
>
> resources/jquery/images/jquery.arrowSteps.head-ltr.png - 303 bytes
> resources/jquery.tipsy/images/tipsy.png - 133 bytes
> resources/jquery.ui/themes/vector/images/ui-icons_ffffff_256x240.png - 3702
> bytes
> resources/mediawiki.action/images/green-checkmark.png - 681 bytes
>
> Why does it matter?  The main reason to avoid requests for every single
> image is to minimize the time, latency, and bandwidth (bytes) used for HTTP
> request and responses.  The smaller the image, the bigger the fraction
> (header size / image size) is, thus the more wasted header time.
>
> What he said about a 814 byte fixed header is also simply wrong.  Some of
> our data URIs are less than 814 bytes *total*.  E.g.
> data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAAkCAAAAABtHw/QAAAAGElEQVQIW2P4wMTA9B8Fo5IIGl0EQQIxAEzaGCA3rKnhAAAAAElFTkSuQmCC
> is a real example from jquery.ui.dialog.
>
> We should keep this issue in mind (there is a penalty, particularly for
> large images), but that article draws a skewed picture.
>
> Matt
>
>
> _______________________________________________
> Design mailing list
> Design@lists.wikimedia.org
> https://lists.wikimedia.org/mailman/listinfo/design



--
Jon Robson
http://jonrobson.me.uk
@rakugojon

_______________________________________________
Design mailing list
Design@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/design